Explainable machine-learning predictions for the prevention of hypoxaemia during surgery

Understanding the mechanism of gully erosion in the alpine region through an interpretable machine learning approach

In the alpine region, climate warming has led to the retreat of glaciers, snow cover, and permafrost. This has intensified water cycling, soil erosion, and increased the occurrence of natural disasters in the alpine region. This study investigated the Lhasa River Basin in the southern Tibetan Plateau, serving as a representative case study of a typical alpine basin, with a specific focus on gully erosion. Based on field investigations and interpretation using high-resolution satellite remote sensing images, the Random Forest (RF) algorithm was applied to evaluate gully erosion susceptibility on watershed level. The Shapley Additive Interpretation method was then used to interpret the RF model and gain deeper insights into the influencing variables of gully erosion. The results showed that the RF model achieved an area under the receiver operating characteristic (AUC) accuracy of 0.99 and 0.98 for the training and testing datasets, respectively, indicating an outstanding performance of the model. The resulting susceptibility map based on the RF model shows that areas with moderate and higher levels of gully erosion susceptibility are covering 50 % of the basin. The model interpretation results indicated that elevation, slope, permafrost, rainstorm, silt loam topsoil, human activity, stream power, and vegetation were the explaining variables with the highest importance for gully erosion occurrence. Different variables are characterized by specific thresholds promoting gully erosion such as: i) elevations higher than 4950 m, ii) slopes steeper than 13.5°, iii) extreme rainstorms longer than 11 days per year, iv) silt loam topsoil, v) presence of permafrost, vi) stream power index higher than 1.2, and vii) normalized difference vegetation index (NDVI) lower than 0.25. Our findings provide the scientific basis to improve soil erosion control in such highly vulnerable alpine area.

PreDBP-PLMs: Prediction of DNA-binding proteins based on pre-trained protein language models and convolutional neural networks

The recognition of DNA-binding proteins (DBPs) is the crucial step to understanding their roles in various biological processes such as genetic regulation, gene expression, cell cycle control, DNA repair, and replication within cells. However, conventional experimental methods for identifying DBPs are usually time-consuming and expensive. Therefore, there is an urgent need to develop rapid and efficient computational methods for the prediction of DBPs. In this study, we proposed a novel predictor named PreDBP-PLMs to further improve the identification accuracy of DBPs by fusing the pre-trained protein language model (PLM) ProtT5 embedding with evolutionary features as input to the classic convolutional neural network (CNN) model. Firstly, the ProtT5 embedding was combined with different evolutionary features derived from the position-specific scoring matrix (PSSM) to represent protein sequences. Then, the optimal feature combination was selected and input to the CNN classifier for the prediction of DBPs. Finally, the 5-fold cross-validation (CV), the leave-one-out CV (LOOCV), and the independent set test were adopted to examine the performance of PreDBP-PLMs on the benchmark datasets. Compared to the existing state-of-the-art predictors, PreDBP-PLMs exhibits an accuracy improvement of 0.5 % and 5.2 % on the PDB186 and PDB2272 datasets, respectively. It demonstrated that the proposed method could serve as a useful tool for the recognition of DBPs.

Danger, Danger, Gaston Labat! Does zero-shot artificial intelligence correlate with anticoagulation guidelines recommendations for neuraxial anesthesia?

INTRODUCTION

Artificial intelligence and large language models (LLMs) have emerged as potentially disruptive technologies in healthcare. In this study GPT-3.5, an accessible LLM, was assessed for its accuracy and reliability in performing guideline-based evaluation of neuraxial bleeding risk in hypothetical patients on anticoagulation medication. The study also explored the impact of structured prompt guidance on the LLM's performance.

METHODS

A dataset of 10 hypothetical patient stems and 26 anticoagulation profiles (260 unique combinations) was developed based on American Society of Regional Anesthesia and Pain Medicine guidelines. Five prompts were created for the LLM, ranging from minimal guidance to explicit instructions. The model's responses were compared with a "truth table" based on the guidelines. Performance metrics, including accuracy and area under the receiver operating curve (AUC), were used.

RESULTS

Baseline performance of GPT-3.5 was slightly above chance. With detailed prompts and explicit guidelines, performance improved significantly (AUC 0.70, 95% CI (0.64 to 0.77)). Performance varied among medication classes.

DISCUSSION

LLMs show potential for assisting in clinical decision making but rely on accurate and relevant prompts. Integration of LLMs should consider safety and privacy concerns. Further research is needed to optimize LLM performance and address complex scenarios. The tested LLM demonstrates potential in assessing neuraxial bleeding risk but relies on precise prompts. LLM integration should be approached cautiously, considering limitations. Future research should focus on optimization and understanding LLM capabilities and limitations in healthcare.

A knowledge-aware deep learning model for landslide susceptibility assessment in Hong Kong

Despite the success of the growing data-driven landslide susceptibility prediction, the model training heavily relies on the quality of the data (involving topography, geology, hydrology, land cover, climate, and human activity), the structure of the model, and the fine-tuning of the model parameters. Few data-driven methods have considered incorporating 'landslide priors', as in this article the prior knowledge or statistics related to landslide occurrence, to enhance the model's perception in landslide mechanism. The main objective and contribution of this study is the coupling of landslide priors and a deep learning model to improve the model's transferability and stability. This is accomplished by selecting non-landslide sample grounded on landslide statistics, disentangling input landslide features using a variational autoencoder, and crafting a loss function with physical constraints. This study utilizes the SHAP method to interpret the deep learning model, aiding in the acquisition of feature permutation results to identify underlying landslide causes. The interpretation result indicates that 'slope' is the most influential factor. Considering the extreme rainfall impact on landslide occurrences in Hong Kong, we combine this prior into the deep learning model and find feature ranking for 'rainfall' improved, in comparison to the ranking result interpreted from a pure MLP. Further, the potency of MT-InSAR is utilized to augment the landslide susceptibility map and promote efficient cross-validation. A comparison of InSAR results with historical images reveals that detectable movement before their occurrence is evident in only a minority of landslides. Most landslides occur spontaneously, exhibiting no precursor motion. Comparing with other data-driven methods, the proposed methods outperform in accuracy (by 2 %-5 %), precision (by 2 %-7 %), recall (by 1 %-3 %), F1-score (by 8 %-10 %), and AuROC (by 2 %-4 %). Especially, the Cohen Kappa performance surpasses nearly 20 %, indicating that the knowledge-aware methodology enhances model generalization and mitigates training bias induced by unbalanced positive and negative samples.

Artificial Intelligence in Metabolomics: A Current Review

Metabolomics and artificial intelligence (AI) form a synergistic partnership. Metabolomics generates large datasets comprising hundreds to thousands of metabolites with complex relationships. AI, aiming to mimic human intelligence through computational modeling, possesses extraordinary capabilities for big data analysis. In this review, we provide a recent overview of the methodologies and applications of AI in metabolomics studies in the context of systems biology and human health. We first introduce the AI concept, history, and key algorithms for machine learning and deep learning, summarizing their strengths and weaknesses. We then discuss studies that have successfully used AI across different aspects of metabolomic analysis, including analytical detection, data preprocessing, biomarker discovery, predictive modeling, and multi-omics data integration. Lastly, we discuss the existing challenges and future perspectives in this rapidly evolving field. Despite limitations and challenges, the combination of metabolomics and AI holds great promises for revolutionary advancements in enhancing human health.

Identification of the best machine learning model for the prediction of driver injury severity

Predicting the injury severities sustained by drivers engaged in road traffic accidents is a key topic of research in road traffic safety. The current study analyzed the driver injury severity (DIS) using twelve machine learning (ML) algorithms. These models were implemented using 0.70, 0.80, and 0.90 train ratios and 5-, 10- and 15-fold cross-validation. Ten years of accident data (from 2011 to 2020) was obtained from police department of Shillong, India. A total of 693 accidents were documented, with 68% being nonfatal and 32% being fatal. Precision, recall, accuracy, F1 score and area under the curve measures were used to compare the performance of all twelve ML models. Overall, the light gradient-boosting machine model was shown to be the best ML model for predicting the injury severities of drivers engaged in road traffic incidents. Finally, variable importance analysis results showed that cause of accident, collision type and types of vehicles were the most influencing factors in nonfatal and fatal driver accidents. The results also revealed that age and gender were slightly associated with DIS. The findings of the current research could be helpful to road safety agencies for the implementation of suitable countermeasures to increase driver safety in road accidents.

Identification of schizophrenia by applying interpretable radiomics modeling with structural magnetic resonance imaging of the cerebellum

AIMS

The cerebellum is involved in higher-order mental processing as well as sensorimotor functions. Although structural abnormalities in the cerebellum have been demonstrated in schizophrenia, neuroimaging techniques are not yet applicable to identify them given the lack of biomarkers. We aimed to develop a robust diagnostic model for schizophrenia using radiomic features from T1-weighted magnetic resonance imaging (T1-MRI) of the cerebellum.

METHODS

A total of 336 participants (174 schizophrenia; 162 healthy controls [HCs]) were allocated to training (122 schizophrenia; 115 HCs) and test (52 schizophrenia; 47 HCs) cohorts. We obtained 2568 radiomic features from T1-MRI of the cerebellar subregions. After feature selection, a light gradient boosting machine classifier was trained. The discrimination and calibration of the model were evaluated. SHapley Additive exPlanations (SHAP) was applied to determine model interpretability.

RESULTS

We identified 17 radiomic features to differentiate participants with schizophrenia from HCs. In the test cohort, the radiomics model had an area under the curve, accuracy, sensitivity, and specificity of 0.89 (95% confidence interval: 0.82-0.95), 78.8%, 88.5%, and 75.4%, respectively. The model explanation by SHAP suggested that the second-order size zone non-uniformity feature from the right lobule IX and first-order energy feature from the right lobules V and VI were highly associated with the risk of schizophrenia.

CONCLUSION

The radiomics model focused on the cerebellum demonstrates robustness in diagnosing schizophrenia. Our results suggest that microcircuit disruption in the posterior cerebellum is a disease-defining feature of schizophrenia, and radiomics modeling has potential for supporting biomarker-based decision-making in clinical practice.

Machine learning-based prediction of diabetic patients using blood routine data

Diabetes stands as one of the most prevalent chronic diseases globally. The conventional methods for diagnosing diabetes are frequently overlooked until individuals manifest noticeable symptoms of the condition. This study aimed to address this gap by collecting comprehensive datasets, including 1000 instances of blood routine data from diabetes patients and an equivalent dataset from healthy individuals. To differentiate diabetes patients from their healthy counterparts, a computational framework was established, encompassing eXtreme Gradient Boosting (XGBoost), random forest, support vector machine, and elastic net algorithms. Notably, the XGBoost model emerged as the most effective, exhibiting superior predictive results with an area under the receiver operating characteristic curve (AUC) of 99.90% in the training set and 98.51% in the testing set. Moreover, the model showcased commendable performance during external validation, achieving an overall accuracy of 81.54%. The probability generated by the model serves as a risk score for diabetes susceptibility. Further interpretability was achieved through the utilization of the Shapley additive explanations (SHAP) algorithm, identifying pivotal indicators such as mean corpuscular hemoglobin concentration (MCHC), lymphocyte ratio (LY%), standard deviation of red blood cell distribution width (RDW-SD), and mean corpuscular hemoglobin (MCH). This enhances our understanding of the predictive mechanisms underlying diabetes. To facilitate the application in clinical and real-life settings, a nomogram was created based on the logistic regression algorithm, which can provide a preliminary assessment of the likelihood of an individual having diabetes. Overall, this research contributes valuable insights into the predictive modeling of diabetes, offering potential applications in clinical practice for more effective and timely diagnoses.

Development of a biomarker prediction model for post-trauma multiple organ failure/dysfunction syndrome based on the blood transcriptome

BACKGROUND

Multiple organ failure/dysfunction syndrome (MOF/MODS) is a major cause of mortality and morbidity among severe trauma patients. Current clinical practices entail monitoring physiological measurements and applying clinical score systems to diagnose its onset. Instead, we aimed to develop an early prediction model for MOF outcome evaluated soon after traumatic injury by performing machine learning analysis of genome-wide transcriptome data from blood samples drawn within 24 h of traumatic injury. We then compared its performance to baseline injury severity scores and detection of infections.

METHODS

Buffy coat transcriptome and linked clinical datasets from blunt trauma patients from the Inflammation and the Host Response to Injury Study ("Glue Grant") multi-center cohort were used. According to the inclusion/exclusion criteria, 141 adult (age ≥ 16 years old) blunt trauma patients (excluding penetrating) with early buffy coat (≤ 24 h since trauma injury) samples were analyzed, with 58 MOF-cases and 83 non-cases. We applied the Least Absolute Shrinkage and Selection Operator (LASSO) and eXtreme Gradient Boosting (XGBoost) algorithms to select features and develop models for MOF early outcome prediction.

RESULTS

The LASSO model included 18 transcripts (AUROC [95% CI]: 0.938 [0.890-0.987] (training) and 0.833 [0.699-0.967] (test)), and the XGBoost model included 41 transcripts (0.999 [0.997-1.000] (training) and 0.907 [0.816-0.998] (test)). There were 16 overlapping transcripts comparing the two panels (0.935 [0.884-0.985] (training) and 0.836 [0.703-0.968] (test)). The biomarker models notably outperformed models based on injury severity scores and sex, which we found to be significantly associated with MOF (APACHEII + sex-0.649 [0.537-0.762] (training) and 0.493 [0.301-0.685] (test); ISS + sex-0.630 [0.516-0.744] (training) and 0.482 [0.293-0.670] (test); NISS + sex-0.651 [0.540-0.763] (training) and 0.525 [0.335-0.714] (test)).

CONCLUSIONS

The accurate assessment of MOF from blood samples immediately after trauma is expected to aid in improving clinical decision-making and may contribute to reduced morbidity, mortality and healthcare costs. Moreover, understanding the molecular mechanisms involving the transcripts identified as important for MOF prediction may eventually aid in developing novel interventions.

Developing a machine learning-based instrument for subjective well-being assessment on Weibo and its psychological significance: An evaluative and interpretive research

Demystifying machine learning (ML) approaches through the synergy of psychology and artificial intelligence can achieve a balance between predictive and explanatory power in model development while enhancing rigor in validation and reporting standards. Accordingly, this study aimed to bridge this research gap by developing a subjective well-being (SWB) prediction model on Weibo, serving as a psychological assessment instrument and explaining the model construction based on psychological knowledge. The model establishment involved the collection of SWB scores and posts from 1,427 valid Weibo users. Multiple machine learning algorithms were employed to train the model and fine-tune its parameters. The optimal model was selected by comparing its criterion validity and split-half reliability performance. Furthermore, SHAP values were calculated to rank the importance of features, which were then used for model interpretation. The criterion validity for the three dimensions of SWB ranged from 0.50 to 0.52 (P < 0.001), and the split-half reliability ranged from 0.94 to 0.96 (P < 0.001). The identified relevant features were related to four main aspects: cultural values, emotions, morality, and time and space. This study expands the application scope of SWB-related psychological theories from a data-driven perspective and provides a theoretical reference for further well-being prediction.

Explore key genes of Crohn's disease based on glycerophospholipid metabolism: A comprehensive analysis Utilizing Mendelian Randomization, Multi-Omics integration, Machine Learning, and SHAP methodology

BACKGROUND AND AIMS

Crohn's disease (CD) is a chronic, complex inflammatory condition with increasing incidence and prevalence worldwide. However, the causes of CD remain incompletely understood. We identified CD-related metabolites, inflammatory factors, and key genes by Mendelian randomization (MR), multi-omics integration, machine learning (ML), and SHAP.

METHODS

We first performed a mediation MR analysis on 1400 serum metabolites, 91 inflammatory factors, and CD. We found that certain phospholipids are causally related to CD. In the scRNA-seq data, monocytes were categorized into high and low metabolism groups based on their glycerophospholipid metabolism scores. The differentially expressed genes of these two groups of cells were extracted, and transcription factor prediction, cell communication analysis, and GSEA analysis were performed. After further screening of differentially expressed genes (FDR<0.05, log2FC>1), least absolute shrinkage and selection operator (LASSO) regression was performed to obtain hub genes. Models for hub genes were built using the Catboost, XGboost, and NGboost methods. Further, we used the SHAP method to interpret the models and obtain the gene with the highest contribution to each model. Finally, qRT-PCR was used to verify the expression of these genes in the peripheral blood mononuclear cells (PBMC) of CD patients and healthy subjects.

RESULT

MR results showed 1-palmitoyl-2-stearoyl-gpc (16:0/18:0) levels, 1-stearoyl-2-arachidonoyl-GPI (18:0/20:4) levels, 1-arachidonoyl-gpc (20:4n6) levels, 1-palmitoyl-2-arachidonoyl-gpc (16:0/20:4n6) levels, and 1-arachidonoyl-GPE (20:4n6) levels were significantly associated with CD risk reduction (FDR<0.05), with CXCL9 acting as a mediation between these phospholipids and CD. The analysis identified 19 hub genes, with Catboost, XGboost, and NGboost achieving AUC of 0.91, 0.88, and 0.85, respectively. The SHAP methodology obtained the three genes with the highest model contribution: G0S2, S100A8, and PLAUR. The qRT-PCR results showed that the expression levels of S100A8 (p = 0.0003), G0S2 (p < 0.0001), and PLAUR (p = 0.0141) in the PBMC of CD patients were higher than healthy subjects.

CONCLUSION

MR findings suggest that certain phospholipids may lower CD risk. G0S2, S100A8, and PLAUR may be potential pathogenic genes in CD. These phospholipids and genes could serve as novel diagnostic and therapeutic targets for CD.

Prediction for Perioperative Stroke Using Intraoperative Parameters

BACKGROUND

Perioperative stroke is a severe complication following surgery. To identify patients at risk for perioperative stroke, several prediction models based on the preoperative factors were suggested. Prediction models often focus on preoperative patient characteristics to assess stroke risk. However, most existing models primarily base their predictions on the patient's baseline characteristics before surgery. We aimed to develop a machine-learning model incorporating both pre- and intraoperative variables to predict perioperative stroke.

METHODS AND RESULTS

This study included patients who underwent noncardiac surgery at 2 hospitals with the data of 15 752 patients from Seoul National University Hospital used for development and temporal internal validation, and the data of 449 patients from Boramae Medical Center used for external validation. Perioperative stroke was defined as a newly developed ischemic lesion on diffusion-weighted imaging within 30 days of surgery. We developed a prediction model composed of pre- and intraoperative factors (integrated model) and compared it with a model consisting of preoperative features alone (preoperative model). Perioperative stroke developed in 109 (0.69%) patients in the Seoul National University Hospital group and 11 patients (2.45%) in the Boramae Medical Center group. The integrated model demonstrated superior predictive performance with area under the curve values of 0.824 (95% CI, 0.762-0.880) versus 0.584 (95% CI, 0.499-0.667; P<0.001) in the internal validation; and 0.716 (95% CI, 0.560-0.859) versus 0.505 (95% CI, 0.343-0.654; P=0.018) in the external validation, compared to the preoperative model.

CONCLUSIONS

We suggest that incorporating intraoperative factors into perioperative stroke prediction models can improve their accuracy.

Multiomics and eXplainable artificial intelligence for decision support in insulin resistance early diagnosis: A pediatric population-based longitudinal study

Pediatric obesity can drastically heighten the risk of cardiometabolic alterations later in life, with insulin resistance standing as the cornerstone linking adiposity to the increased cardiovascular risk. Puberty has been pointed out as a critical stage after which obesity-associated insulin resistance is more difficult to revert. Timely prediction of insulin resistance in pediatric obesity is therefore vital for mitigating the risk of its associated comorbidities. The construction of effective and robust predictive systems for a complex health outcome like insulin resistance during the early stages of life demands the adoption of longitudinal designs for more causal inferences, and the integration of factors of varying nature involved in its onset. In this work, we propose an eXplainable Artificial Intelligence-based decision support pipeline for early diagnosis of insulin resistance in a longitudinal cohort of 90 children. For that, we leverage multi-omics (genomics and epigenomics) and clinical data from the pre-pubertal stage. Different data layers combinations, pre-processing techniques (missing values, feature selection, class imbalance, etc.), algorithms, training procedures were considered following good practices for Machine Learning. SHapley Additive exPlanations were provided for specialists to understand both the decision-making mechanisms of the system and the impact of the features on each automatic decision, an essential issue in high-risk areas such as this one where system decisions may affect people's lives. The system showed a relevant predictive ability (AUC and G-mean of 0.92). A deep exploration, both at the global and the local level, revealed promising biomarkers of insulin resistance in our population, highlighting classical markers, such as Body Mass Index z-score or leptin/adiponectin ratio, and novel ones such as methylation patterns of relevant genes, such as HDAC4, PTPRN2, MATN2, RASGRF1 and EBF1. Our findings highlight the importance of integrating multi-omics data and following eXplainable Artificial Intelligence trends when building decision support systems.

Machine learning-derived dose-response relationships considering interactions in mixtures: Applications to the oxidative potential of particulate matter

Conventional environmental health research is primarily focused on isolated chemical exposures, neglecting the complex interactions between multiple pollutants that may synergistically or antagonistically influence toxicity, thereby posing unexpected health risks. In this study, we address this knowledge gap by introducing an explainable machine learning (ML) approach with Feature Localized Intercept Transformed-Shapley Additive Explanations (FLIT-SHAP) designed to extract the dose-response relationships of specific pollutants in mixtures. In contrast to traditional SHAP, FLIT-SHAP can localize the global intercept to elucidate mixture effects, which is crucial for understanding the oxidative potential (OP) of ambient particulate matter (PM). Assessing multi-pollutant OP using FLIT-SHAP revealed both synergistic (55-63 %) and antagonistic (25-42 %) effects in laboratory-controlled OP data, but an antagonistic (33-66 %; lower OP) effect in ambient PM. Notably, the FLIT-SHAP approach demonstrated higher prediction accuracy (R2 = 0.99) compared to the additive model (R2 = 0.89) when evaluated against real-world PM samples. Quinones, such as phenanthrenequinone, play a more significant role in PM2.5 than previously recognized. Through this study, we highlighted the potential of FLIT-SHAP to enhance toxicity predictions and aid decision-making in the field of environmental health.

Artificial intelligence-based predictive model for guidance on treatment strategy selection in oral and maxillofacial surgery

Application of deep learning (DL) and machine learning (ML) is rapidly increasing in the medical field. DL is gaining significance for medical image analysis, particularly, in oral and maxillofacial surgeries. Owing to the ability to accurately identify and categorize both diseased and normal soft- and hard-tissue structures, DL has high application potential in the diagnosis and treatment of tumors and in orthognathic surgeries. Moreover, DL and ML can be used to develop prediction models that can aid surgeons to assess prognosis by analyzing the patient's medical history, imaging data, and surgical records, develop more effective treatment strategies, select appropriate surgical modalities, and evaluate the risk of postoperative complications. Such prediction models can play a crucial role in the selection of treatment strategies for oral and maxillofacial surgeries. Their practical application can improve the utilization of medical staff, increase the treatment accuracy and efficiency, reduce surgical risks, and provide an enhanced treatment experience to patients. However, DL and ML face limitations, such as data drift, unstable model results, and vulnerable social trust. With the advancement of social concepts and technologies, the use of these models in oral and maxillofacial surgery is anticipated to become more comprehensive and extensive.

Enhancing metagenomic classification with compression-based features

Metagenomics is a rapidly expanding field that uses next-generation sequencing technology to analyze the genetic makeup of environmental samples. However, accurately identifying the organisms in a metagenomic sample can be complex, and traditional reference-based methods may need to be more effective in some instances. In this study, we present a novel approach for metagenomic identification, using data compressors as a feature for taxonomic classification. By evaluating a comprehensive set of compressors, including both general-purpose and genomic-specific, we demonstrate the effectiveness of this method in accurately identifying organisms in metagenomic samples. The results indicate that using features from multiple compressors can help identify taxonomy. An overall accuracy of 95% was achieved using this method using an imbalanced dataset with classes with limited samples. The study also showed that the correlation between compression and classification is insignificant, highlighting the need for a multi-faceted approach to metagenomic identification. This approach offers a significant advancement in the field of metagenomics, providing a reference-less method for taxonomic identification that is both effective and efficient while revealing insights into the statistical and algorithmic nature of genomic data. The code to validate this study is publicly available at https://github.com/ieeta-pt/xgTaxonomy.

Exploring high-emission driving behaviors of heavy-duty diesel vehicles based on engine principles under different road grade levels

To reveal the outstanding high-emission problems that occur when heavy-duty diesel vehicles (HDDV) pass uphill and downhill, this study proposes a method to depict the nitrogen oxides (NOx) and carbon dioxide (CO2) high-emission driving behaviors caused by slopes from the perspective of engine principles. By calculating emission and grade data of HDDV based on on-board diagnostic (OBD) data and digital elevation model (DEM) data, the 262 short trips including uphill, flat-road and downhill are firstly obtained through the rule-based short trip segmentation method, and the significant correlation between the road grade and emissions of the short trips is verified by Kendall's Tau and K-means clustering. Secondly, by comparing the distribution changes of three speed categories (acceleration state, constant speed state and deceleration state), the differences in HDDV operating states under different grade levels are discussed. Finally, the machine learning models (Random Forest, XGBoost and Elastic Net), are used to develop the NOx and CO2 emission estimation model, identifying high-emission driving behaviors, particularly during uphill driving, which showed the highest proportion of high-emission. Explained by the feature importance and SHapley Additive exPlanations (SHAP) model that large accelerator pedal opening, frequent aggressive acceleration, and high engine load have positive effects both on NOx and CO2 emissions. The difference is in the air-fuel ratio that the engine in the rich or slightly lean burning state will increase CO2 emissions and the lean burning state will increase NOx emissions. In addition, due to the uncertainty of the actual uphill, drivers often undergo a rapid "deceleration-uniform-acceleration" process, which significantly contributes to high NOx and CO2 emissions from the engine perspective. The findings provide insights for designing driving strategies in slope scenarios and offer a novel perspective on depicting driving behaviors.

Factors affecting biochemical pregnancy loss (BPL) in preimplantation genetic testing for aneuploidy (PGT-A) cycles: machine learning-assisted identification

PURPOSE

To determine the factors influencing the likelihood of biochemical pregnancy loss (BPL) after transfer of a euploid embryo from preimplantation genetic testing for aneuploidy (PGT-A) cycles.

METHODS

The study employed an observational, retrospective cohort design, encompassing 6020 embryos from 2879 PGT-A cycles conducted between February 2013 and September 2021. Trophectoderm biopsies in day 5 (D5) or day 6 (D6) blastocysts were analyzed by next generation sequencing (NGS). Only single embryo transfers (SET) were considered, totaling 1161 transfers. Of these, 49.9% resulted in positive pregnancy tests, with 18.3% experiencing BPL. To establish a predictive model for BPL, both classical statistical methods and five different supervised classification machine learning algorithms were used. A total of forty-seven factors were incorporated as predictor variables in the machine learning models.

RESULTS

Throughout the optimization process for each model, various performance metrics were computed. Random Forest model emerged as the best model, boasting the highest area under the ROC curve (AUC) value of 0.913, alongside an accuracy of 0.830, positive predictive value of 0.857, and negative predictive value of 0.807. For the selected model, SHAP (SHapley Additive exPlanations) values were determined for each of the variables to establish which had the best predictive ability. Notably, variables pertaining to embryo biopsy demonstrated the greatest predictive capacity, followed by factors associated with ovarian stimulation (COS), maternal age, and paternal age.

CONCLUSIONS

The Random Forest model had a higher predictive power for identifying BPL occurrences in PGT-A cycles. Specifically, variables associated with the embryo biopsy procedure (biopsy day, number of biopsied embryos, and number of biopsied cells) and ovarian stimulation (number of oocytes retrieved and duration of stimulation), exhibited the strongest predictive power.

Proteomics and machine learning in the prediction and explanation of low pectoralis muscle area

Low muscle mass is associated with numerous adverse outcomes independent of other associated comorbid diseases. We aimed to predict and understand an individual's risk for developing low muscle mass using proteomics and machine learning. We identified eight biomarkers associated with low pectoralis muscle area (PMA). We built three random forest classification models that used either clinical measures, feature selected biomarkers, or both to predict development of low PMA. The area under the receiver operating characteristic curve for each model was: clinical-only = 0.646, biomarker-only = 0.740, and combined = 0.744. We displayed the heterogenetic nature of an individual's risk for developing low PMA and identified two distinct subtypes of participants who developed low PMA. While additional validation is required, our methods for identifying and understanding individual and group risk for low muscle mass could be used to enable developments in the personalized prevention of low muscle mass.

The Emerging and Important Role of Artificial Intelligence in Cardiac Surgery

Artificial Intelligence (AI) has greatly affected our everyday lives and holds great promise to change the landscape of medicine. AI is particularly positioned to improve care for the increasingly complex patients undergoing cardiac surgery utilizing immense amount of data generated in the course of their care. When deployed, AI can be used to analyze this information at the patient's bedside more expediently and accurately, all while providing new insights. This review summarizes the current applications of AI in cardiac surgery, from the vantage point of a patient's journey. Applications of AI include pre-operative risk assessment, intraoperative planning, post-operative patient care and out-patient telemonitoring, encompassing the spectrum of cardiac surgical care. Offloading of administrative processes and enhanced experience with information gathering also represent a unique and underrepresented avenue for future utilization of AI. As clinicians, understanding the nomenclature and applications of AI is important to contextualize problems, to ensure problem-driven solutions and for clinical benefit. Precision medicine, and thus clinically relevant AI, remains dependent on data curation and warehousing to gather insights from large multicenter repositories while treating privacy with the utmost importance. AI tasks should not be siloed but rather holistically integrated into clinical workflow to retain context and relevance. As cardiac surgeons, AI allows us to look forward to a bright future of more efficient utilization of our clinical expertise toward high-level decision making and technical prowess.

Is the coupling and coordination of economic, social and environmental development crucial to the governance of relative poverty?

As the primary goal of the 17 Sustainable Development Goals (SDGs), poverty eradication is still one of the major challenges faced by countries around the world, and relative poverty is a comprehensive poverty pattern triggered by the superposition of economic, social, and environmental dimensions. Therefore, Therefore, this paper introduces the perspective of coupled coordination to consider the formation of relative poverty, constructs indicators in three major dimensions: economic, social, and environmental, proposes a fast and more accurate method of identifying relative poverty in a region by using machine learning, measures the degree of coupled coordination of China's relatively poor provinces using a coupled coordination model and analyzes the relationship with the level of relative poverty, and puts forward suggestions for poverty management on this basis using typology classification. The results of the study show that: 1) the fusion of data crawlers, remote sensing space, and other multi-source data to construct the dataset and propose a fast and efficient regional relative poverty identification method based on big data with low comprehensive cost and high identification accuracy of 0.914. 2) Currently, 70.83% of the economic-social-environmental systems of the relatively poor regions are in the dysfunctional type and are in a state of disordered development and malignant constraints. The regions showing coupling disorders are mainly clustered in the three southern prefectures of Xinjiang, Qinghai, Gansu, Yunnan, and Sichuan, and their spatial distribution is relatively concentrated. 3) The types of poverty and their coupled and coordinated development in each region show large spatial variability, requiring differentiated poverty eradication countermeasures tailored to local conditions to achieve sustainable regional economic-social-environmental development.

Interpretable artificial intelligence to optimise use of imatinib after resection in patients with localised gastrointestinal stromal tumours: an observational cohort study

BACKGROUND

Current guidelines recommend use of adjuvant imatinib therapy for many patients with gastrointestinal stromal tumours (GISTs); however, its optimal treatment duration is unknown and some patient groups do not benefit from the therapy. We aimed to apply state-of-the-art, interpretable artificial intelligence (ie, predictions or prescription logic that can be easily understood) methods on real-world data to establish which groups of patients with GISTs should receive adjuvant imatinib, its optimal treatment duration, and the benefits conferred by this therapy.

METHODS

In this observational cohort study, we considered for inclusion all patients who underwent resection of primary, non-metastatic GISTs at the Memorial Sloan Kettering Cancer Center (MSKCC; New York, NY, USA) between Oct 1, 1982, and Dec 31, 2017, and who were classified as intermediate or high risk according to the Armed Forces Institute of Pathology Miettinen criteria and had complete follow-up data with no missing entries. A counterfactual random forest model, which used predictors of recurrence (mitotic count, tumour size, and tumour site) and imatinib duration to infer the probability of recurrence at 7 years for a given patient under each duration of imatinib treatment, was trained in the MSKCC cohort. Optimal policy trees (OPTs), a state-of-the-art interpretable AI-based method, were used to read the counterfactual random forest model by training a decision tree with the counterfactual predictions. The OPT recommendations were externally validated in two cohorts of patients from Poland (the Polish Clinical GIST Registry), who underwent GIST resection between Dec 1, 1981, and Dec 31, 2011, and from Spain (the Spanish Group for Research in Sarcomas), who underwent resection between Oct 1, 1987, and Jan 30, 2011.

FINDINGS

Among 1007 patients who underwent GIST surgery in MSKCC, 117 were included in the internal cohort; for the external cohorts, the Polish cohort comprised 363 patients and the Spanish cohort comprised 239 patients. The OPT did not recommend imatinib for patients with GISTs of gastric origin measuring less than 15·9 cm with a mitotic count of less than 11·5 mitoses per 5 mm2 or for those with small GISTs (<5·4 cm) of any site with a count of less than 11·5 mitoses per 5 mm2. In this cohort, the OPT cutoffs had a sensitivity of 92·7% (95% CI 82·4-98·0) and a specificity of 33·9% (22·3-47·0). The application of these cutoffs in the two external cohorts would have spared 38 (29%) of 131 patients in the Spanish cohort and 44 (35%) of 126 patients in the Polish cohort from unnecessary treatment with imatinib. Meanwhile, the risk of undertreating patients in these cohorts was minimal (sensitivity 95·4% [95% CI 89·5-98·5] in the Spanish cohort and 92·4% [88·3-95·4] in the Polish cohort). The OPT tested 33 different durations of imatinib treatment (<5 years) and found that 5 years of treatment conferred the most benefit.

INTERPRETATION

If the identified patient subgroups were applied in clinical practice, as many as a third of the current cohort of candidates who do not benefit from adjuvant imatinib would be encouraged to not receive imatinib, subsequently avoiding unnecessary toxicity on patients and financial strain on health-care systems. Our finding that 5 years is the optimal duration of imatinib treatment could be the best source of evidence to inform clinical practice until 2028, when a randomised controlled trial with the same aims is expected to report its findings.

FUNDING

National Cancer Institute.

Establishment and validation of an artificial intelligence web application for predicting postoperative in-hospital mortality in patients with hip fracture: a national cohort study of 52 707 cases

BACKGROUND

In-hospital mortality following hip fractures is a significant concern, and accurate prediction of this outcome is crucial for appropriate clinical management. Nonetheless, there is a lack of effective prediction tools in clinical practice. By utilizing artificial intelligence (AI) and machine learning techniques, this study aims to develop a predictive model that can assist clinicians in identifying geriatric hip fracture patients at a higher risk of in-hospital mortality.

METHODS

A total of 52 707 geriatric hip fracture patients treated with surgery from 90 hospitals were included in this study. The primary outcome was postoperative in-hospital mortality. The patients were randomly divided into two groups, with a ratio of 7:3. The majority of patients, assigned to the training cohort, were used to develop the AI models. The remaining patients, assigned to the validation cohort, were used to validate the models. Various machine learning algorithms, including logistic regression (LR), decision tree (DT), naïve bayesian (NB), neural network (NN), eXGBoosting machine (eXGBM), and random forest (RF), were employed for model development. A comprehensive scoring system, incorporating 10 evaluation metrics, was developed to assess the prediction performance, with higher scores indicating superior predictive capability. Based on the best machine learning-based model, an AI application was developed on the Internet. In addition, a comparative testing of prediction performance between doctors and the AI application.

FINDINGS

The eXGBM model exhibited the best prediction performance, with an area under the curve (AUC) of 0.908 (95% CI: 0.881-0.932), as well as the highest accuracy (0.820), precision (0.817), specificity (0.814), and F1 score (0.822), and the lowest Brier score (0.120) and log loss (0.374). Additionally, the model showed favorable calibration, with a slope of 0.999 and an intercept of 0.028. According to the scoring system incorporating 10 evaluation metrics, the eXGBM model achieved the highest score (56), followed by the RF model (48) and NN model (41). The LR, DT, and NB models had total scores of 27, 30, and 13, respectively. The AI application has been deployed online at https://in-hospitaldeathinhipfracture-l9vhqo3l55fy8dkdvuskvu.streamlit.app/ , based on the eXGBM model. The comparative testing revealed that the AI application's predictive capabilities significantly outperformed those of the doctors in terms of AUC values (0.908 vs. 0.682, P <0.001).

CONCLUSIONS

The eXGBM model demonstrates promising predictive performance in assessing the risk of postoperative in-hospital mortality among geriatric hip fracture patients. The developed AI model serves as a valuable tool to enhance clinical decision-making.

Artificial intelligence and nonoperating room anesthesia

PURPOSE OF REVIEW

The integration of artificial intelligence (AI) in nonoperating room anesthesia (NORA) represents a timely and significant advancement. As the demand for NORA services expands, the application of AI is poised to improve patient selection, perioperative care, and anesthesia delivery. This review examines AI's growing impact on NORA and how it can optimize our clinical practice in the near future.

RECENT FINDINGS

AI has already improved various aspects of anesthesia, including preoperative assessment, intraoperative management, and postoperative care. Studies highlight AI's role in patient risk stratification, real-time decision support, and predictive modeling for patient outcomes. Notably, AI applications can be used to target patients at risk of complications, alert clinicians to the upcoming occurrence of an intraoperative adverse event such as hypotension or hypoxemia, or predict their tolerance of anesthesia after the procedure. Despite these advances, challenges persist, including ethical considerations, algorithmic bias, data security, and the need for transparent decision-making processes within AI systems.

SUMMARY

The findings underscore the substantial benefits of AI in NORA, which include improved safety, efficiency, and personalized care. AI's predictive capabilities in assessing hypoxemia risk and other perioperative events, have demonstrated potential to exceed human prognostic accuracy. The implications of these findings advocate for a careful yet progressive adoption of AI in clinical practice, encouraging the development of robust ethical guidelines, continual professional training, and comprehensive data management strategies. Furthermore, AI's role in anesthesia underscores the need for multidisciplinary research to address the limitations and fully leverage AI's capabilities for patient-centered anesthesia care.

Improving the performance of deep learning models in predicting and classifying gamma passing rates with discriminative features and a class balancing technique: a retrospective cohort study

BACKGROUND

The purpose of this study was to improve the deep learning (DL) model performance in predicting and classifying IMRT gamma passing rate (GPR) by using input features related to machine parameters and a class balancing technique.

METHODS

A total of 2348 fields from 204 IMRT plans for patients with nasopharyngeal carcinoma were retrospectively collected to form a dataset. Input feature maps, including fluence, leaf gap, leaf speed of both banks, and corresponding errors, were constructed from the dynamic log files. The SHAP framework was employed to compute the impact of each feature on the model output for recursive feature elimination. A series of UNet++ based models were trained on the obtained eight feature sets with three fine-tuning methods including the standard mean squared error (MSE) loss, a re-sampling technique, and a proposed weighted MSE loss (WMSE). Differences in mean absolute error, area under the receiver operating characteristic curve (AUC), sensitivity, and specificity were compared between the different models.

RESULTS

The models trained with feature sets including leaf speed and leaf gap features predicted GPR for failed fields more accurately than the other models (F(7, 147) = 5.378, p < 0.001). The WMSE loss had the highest accuracy in predicting GPR for failed fields among the three fine-tuning methods (F(2, 42) = 14.149, p < 0.001), while an opposite trend was observed in predicting GPR for passed fields (F(2, 730) = 9.907, p < 0.001). The WMSE_FS5 model achieved a superior AUC (0.92) and more balanced sensitivity (0.77) and specificity (0.89) compared to the other models.

CONCLUSIONS

Machine parameters can provide discriminative input features for GPR prediction in DL. The novel weighted loss function demonstrates the ability to balance the prediction and classification accuracy between the passed and failed fields. The proposed approach is able to improve the DL model performance in predicting and classifying GPR, and can potentially be integrated into the plan optimization process to generate higher deliverability plans.

TRIAL REGISTRATION

This clinical trial was registered in the Chinese Clinical Trial Registry on March 26th, 2020 (registration number: ChiCTR2000031276). https://clinicaltrials.gov/ct2/show/ChiCTR2000031276.

A Machine Learning Model for Predicting In-Hospital Mortality in Chinese Patients With ST-Segment Elevation Myocardial Infarction: Findings From the China Myocardial Infarction Registry

BACKGROUND

Machine learning (ML) risk prediction models, although much more accurate than traditional statistical methods, are inconvenient to use in clinical practice due to their nontransparency and requirement of a large number of input variables.

OBJECTIVE

We aimed to develop a precise, explainable, and flexible ML model to predict the risk of in-hospital mortality in patients with ST-segment elevation myocardial infarction (STEMI).

METHODS

This study recruited 18,744 patients enrolled in the 2013 China Acute Myocardial Infarction (CAMI) registry and 12,018 patients from the China Patient-Centered Evaluative Assessment of Cardiac Events (PEACE)-Retrospective Acute Myocardial Infarction Study. The Extreme Gradient Boosting (XGBoost) model was derived from 9616 patients in the CAMI registry (2014, 89 variables) with 5-fold cross-validation and validated on both the 9125 patients in the CAMI registry (89 variables) and the independent China PEACE cohort (10 variables). The Shapley Additive Explanations (SHAP) approach was employed to interpret the complex relationships embedded in the proposed model.

RESULTS

In the XGBoost model for predicting all-cause in-hospital mortality, the variables with the top 8 most important scores were age, left ventricular ejection fraction, Killip class, heart rate, creatinine, blood glucose, white blood cell count, and use of angiotensin-converting enzyme inhibitors (ACEIs) and angiotensin II receptor blockers (ARBs). The area under the curve (AUC) on the CAMI validation set was 0.896 (95% CI 0.884-0.909), significantly higher than the previous models. The AUC for the Global Registry of Acute Coronary Events (GRACE) model was 0.809 (95% CI 0.790-0.828), and for the TIMI model, it was 0.782 (95% CI 0.763-0.800). Despite the China PEACE validation set only having 10 available variables, the AUC reached 0.840 (0.829-0.852), showing a substantial improvement to the GRACE (0.762, 95% CI 0.748-0.776) and TIMI (0.789, 95% CI 0.776-0.803) scores. Several novel and nonlinear relationships were discovered between patients' characteristics and in-hospital mortality, including a U-shape pattern of high-density lipoprotein cholesterol (HDL-C).

CONCLUSIONS

The proposed ML risk prediction model was highly accurate in predicting in-hospital mortality. Its flexible and explainable characteristics make the model convenient to use in clinical practice and could help guide patient management.

TRIAL REGISTRATION

ClinicalTrials.gov NCT01874691; https://clinicaltrials.gov/study/NCT01874691.

Effects of Bony Pelvic and Prostate Dimensions on Surgical Difficulty of Robot-Assisted Radical Prostatectomy: An Original Study and Meta-analysis

BACKGROUND

Due to the deep location of the prostate within the pelvic cavity, procedures of robot-assisted radical prostatectomy (RARP) might be challenged by the prostate size and the limited pelvic cavity space. This study aimed to investigate the roles of bony pelvic and prostate dimensions in RARP procedures by an original study coupled with a meta-analysis.

METHODS

In the original study, patients undergoing multiport RARP between 2021 and 2022 were consecutively assessed. The associations of anatomic features with operative time (OT), estimated blood loss (EBL), and positive surgical margin (PSM) were evaluated using linear and logistic regression analyses as well as restricted cubic spline (RCS) analysis. Based on machine-learning algorithms, this study established predictive models for surgical difficulty and interpreted the model using SHapley Additive exPlanation (SHAP). In the meta-analysis, three databases were searched for eligible studies. Quantitative syntheses were subsequently performed.

RESULTS

Overall, 219 patients were enrolled in the original study. Prostate volume (PV) and the prostate volume-to-pelvic cavity index (PCI) ratio (PV-to-PCI ratio) were significantly associated with longer OT (P < 0.05). In the RCS models, U-shaped associations were observed between the prostate anteroposterior diameter (PAD) and OT, and between the prostate height (PH) and EBL, and an L-shaped association was observed between the anteroposterior diameter of the pelvic inlet (API) and EBL. The XGBoost model was superior to the logistic regression model in predicting prolonged OT. The meta-analysis demonstrated that greater PV was significantly associated with longer OT (β = 0.20; 95% confidence interval [CI] 0.12-0.27; odds ratio [OR] = 1.05; 95% CI 1.00-1.11), and a smaller PV could increase the risk of PSM (OR = 0.82; 95% CI 0.77-0.88).

CONCLUSIONS

A large prostate within a narrow and deep pelvis might suggest increased surgical difficulty of RARP. The size of the pelvic inlet also had a great impact on RARP. For PAD and PH, there seemed to be an optimal range with the lowest surgical difficulty. Machine-learning models based on the XGBoost algorithm could be successfully applied to predict the surgical difficulty of RARP.

Off-Label use of Woven EndoBridge device for intracranial brain aneurysm treatment: Modeling of occlusion outcome

INTRODUCTION

The Woven EndoBridge (WEB) device is emerging as a novel therapy for intracranial aneurysms, but its use for off-label indications requires further study. Using machine learning, we aimed to develop predictive models for complete occlusion after off-label WEB treatment and to identify factors associated with occlusion outcomes.

METHODS

This multicenter, retrospective study included 162 patients who underwent off-label WEB treatment for intracranial aneurysms. Baseline, morphological, and procedural variables were utilized to develop machine-learning models predicting complete occlusion. Model interpretation was performed to determine significant predictors. Ordinal regression was also performed with occlusion status as an ordinal outcome from better (Raymond Roy Occlusion Classification [RROC] grade 1) to worse (RROC grade 3) status. Odds ratios (OR) with 95 % confidence intervals (CI) were reported.

RESULTS

The best performing model achieved an AUROC of 0.8 for predicting complete occlusion. Larger neck diameter and daughter sac were significant independent predictors of incomplete occlusion. On multivariable ordinal regression, higher RROC grades (OR 1.86, 95 % CI 1.25-2.82), larger neck diameter (OR 1.69, 95 % CI 1.09-2.65), and presence of daughter sacs (OR 2.26, 95 % CI 0.99-5.15) were associated with worse aneurysm occlusion after WEB treatment, independent of other factors.

CONCLUSION

This study found that larger neck diameter and daughter sacs were associated with worse occlusion after WEB therapy for aneurysms. The machine learning approach identified anatomical factors related to occlusion outcomes that may help guide patient selection and monitoring with this technology. Further validation is needed.

Machine learning models including preoperative and postoperative albumin-bilirubin score: short-term outcomes among patients with hepatocellular carcinoma

BACKGROUND

We sought to assess the impact of various perioperative factors on the risk of severe complications and post-surgical mortality using a novel maching learning technique.

METHODS

Data on patients undergoing resection for HCC were obtained from an international, multi-institutional database between 2000 and 2020. Gradient boosted trees were utilized to construct predictive models.

RESULTS

Among 962 patients who underwent HCC resection, the incidence of severe postoperative complications was 12.7% (n = 122); in-hospital mortality was 2.9% (n = 28). Models that exclusively used preoperative data achieved AUC values of 0.89 (95%CI 0.85 to 0.92) and 0.90 (95%CI 0.84 to 0.96) to predict severe complications and mortality, respectively. Models that combined preoperative and postoperative data achieved AUC values of 0.93 (95%CI 0.91 to 0.96) and 0.92 (95%CI 0.86 to 0.97) for severe morbidity and mortality, respectively. The SHAP algorithm demonstrated that the factor most strongly predictive of severe morbidity and mortality was postoperative day 1 and 3 albumin-bilirubin (ALBI) scores.

CONCLUSION

Incorporation of perioperative data including ALBI scores using ML techniques can help risk-stratify patients undergoing resection of HCC.

Explainable AI decision support improves accuracy during telehealth strep throat screening

BACKGROUND

Artificial intelligence-based (AI) clinical decision support systems (CDSS) using unconventional data, like smartphone-acquired images, promise transformational opportunities for telehealth; including remote diagnosis. Although such solutions' potential remains largely untapped, providers' trust and understanding are vital for effective adoption. This study examines how different human-AI interaction paradigms affect clinicians' responses to an emerging AI CDSS for streptococcal pharyngitis (strep throat) detection from smartphone throat images.

METHODS

In a randomized experiment, we tested explainable AI strategies using three AI-based CDSS prototypes for strep throat prediction. Participants received clinical vignettes via an online survey to predict the disease state and offer clinical recommendations. The first set included a validated CDSS prediction (Modified Centor Score) and the second introduced an explainable AI prototype randomly. We used linear models to assess explainable AI's effect on clinicians' accuracy, confirmatory testing rates, and perceived trust and understanding of the CDSS.

RESULTS

The study, involving 121 telehealth providers, shows that compared to using the Centor Score, AI-based CDSS can improve clinicians' predictions. Despite higher agreement with AI, participants report lower trust in its advice than in the Centor Score, leading to more requests for in-person confirmatory testing.

CONCLUSIONS

Effectively integrating AI is crucial in the telehealth-based diagnosis of infectious diseases, given the implications of antibiotic over-prescriptions. We demonstrate that AI-based CDSS can improve the accuracy of remote strep throat screening yet underscores the necessity to enhance human-machine collaboration, particularly in trust and intelligibility. This ensures providers and patients can capitalize on AI interventions and smartphones for virtual healthcare.

Boosting the Accuracy and Chemical Space Coverage of the Detection of Small Colloidal Aggregating Molecules Using the BAD Molecule Filter

The ability to conduct effective high throughput screening (HTS) campaigns in drug discovery is often hampered by the detection of false positives in these assays due to small colloidally aggregating molecules (SCAMs). SCAMs can produce artifactual hits in HTS by nonspecific inhibition of the protein target. In this work, we present a new computational prediction tool for detecting SCAMs based on their 2D chemical structure. The tool, called the boosted aggregation detection (BAD) molecule filter, employs decision tree ensemble methods, namely, the CatBoost classifier and the light gradient-boosting machine, to significantly improve the detection of SCAMs. In developing the filter, we explore models trained on individual data sets, a consensus approach using these models, and, third, a merged data set approach, each tailored for specific drug discovery needs. The individual data set method emerged as most effective, achieving 93% sensitivity and 90% specificity, outperforming existing state-of-the-art models by 20 and 5%, respectively. The consensus models offer broader chemical space coverage, exceeding 90% for all testing sets. This feature is an important aspect particularly for early stage medicinal chemistry projects, and provides information on applicability domain. Meanwhile, the merged data set models demonstrated robust performance, with a notable sensitivity of 79% in the comprehensive 10-fold cross-validation test set. A SHAP analysis of model features indicates the importance of hydrophobicity and molecular complexity as primary factors influencing the aggregation propensity. The BAD molecule filter is readily accessible for the public usage on https://molmodlab-aau.com/Tools.html. This filter provides a new, more robust tool for aggregate prediction in the early stages of drug discovery to optimize hit rates and reduce associated testing and validation overheads.

Development and validation of machine learning models to predict MDRO colonization or infection on ICU admission by using electronic health record data

BACKGROUND

Multidrug-resistant organisms (MDRO) pose a significant threat to public health. Intensive Care Units (ICU), characterized by the extensive use of antimicrobial agents and a high prevalence of bacterial resistance, are hotspots for MDRO proliferation. Timely identification of patients at high risk for MDRO can aid in curbing transmission, enhancing patient outcomes, and maintaining the cleanliness of the ICU environment. This study focused on developing a machine learning (ML) model to identify patients at risk of MDRO during the initial phase of their ICU stay.

METHODS

Utilizing patient data from the First Medical Center of the People's Liberation Army General Hospital (PLAGH-ICU) and the Medical Information Mart for Intensive Care (MIMIC-IV), the study analyzed variables within 24 h of ICU admission. Machine learning algorithms were applied to these datasets, emphasizing the early detection of MDRO colonization or infection. Model efficacy was evaluated by the area under the receiver operating characteristics curve (AUROC), alongside internal and external validation sets.

RESULTS

The study evaluated 3,536 patients in PLAGH-ICU and 34,923 in MIMIC-IV, revealing MDRO prevalence of 11.96% and 8.81%, respectively. Significant differences in ICU and hospital stays, along with mortality rates, were observed between MDRO positive and negative patients. In the temporal validation, the PLAGH-ICU model achieved an AUROC of 0.786 [0.748, 0.825], while the MIMIC-IV model reached 0.744 [0.723, 0.766]. External validation demonstrated reduced model performance across different datasets. Key predictors included biochemical markers and the duration of pre-ICU hospital stay.

CONCLUSIONS

The ML models developed in this study demonstrated their capability in early identification of MDRO risks in ICU patients. Continuous refinement and validation in varied clinical contexts remain essential for future applications.

Protocol for performing metabolic pathway-based subtyping of breast tumors

Here, we present a protocol for analyzing the global metabolic landscape in breast tumors for the purpose of metabolism-based patient stratification. We describe steps for analyzing 1,454 metabolic genes representing 90 metabolic pathways and subjecting them to an algorithm that calculates the deregulation score of 90 pathways in each tumor sample, thus converting gene-level information into pathway-level information. We then detail procedures for performing clustering analysis to identify metabolic subtypes and using machine learning to develop a signature representing each subtype. For complete details on the use and execution of this protocol, please refer to Iqbal et al.1.

Deep Learning-Based Dynamic Risk Prediction of Venous Thromboembolism for Patients With Ovarian Cancer in Real-World Settings From Electronic Health Records

PURPOSE

Patients with epithelial ovarian cancer (EOC) have an elevated risk for venous thromboembolism (VTE). To assess the risk of VTE, models were developed by statistical or machine learning algorithms. However, few models have accommodated deep learning (DL) algorithms in realistic clinical settings. We aimed to develop a predictive DL model, exploiting rich information from electronic health records (EHRs), including dynamic clinical features and the presence of competing risks.

METHODS

We extracted EHRs of 1,268 patients diagnosed with EOC from January 2007 through December 2017 at the National Cancer Center, Korea. DL survival networks using fully connected layers, temporal attention, and recurrent neural networks were adopted and compared with multi-perceptron-based classification models. Prediction accuracy was independently validated in the data set of 423 patients newly diagnosed with EOC from January 2018 to December 2019. Personalized risk plots displaying the individual interval risk were developed.

RESULTS

DL-based survival networks achieved a superior area under the receiver operating characteristic curve (AUROC) between 0.95 and 0.98 while the AUROC of classification models was between 0.85 and 0.90. As clinical information benefits the prediction accuracy, the proposed dynamic survival network outperformed other survival networks for the test and validation data set with the highest time-dependent concordance index (0.974, 0.975) and lowest Brier score (0.051, 0.049) at 6 months after a cancer diagnosis. Our visualization showed that the interval risk fluctuating along with the changes in longitudinal clinical features.

CONCLUSION

Adaption of dynamic patient clinical features and accounting for competing risks from EHRs into the DL algorithms demonstrated VTE risk prediction with high accuracy. Our results show that this novel dynamic survival network can provide personalized risk prediction with the potential to assist risk-based clinical intervention to prevent VTE among patients with EOC.

Deep learning prediction and experimental investigation of specific capacitance of nitrogen-doped porous biochar

N-doped porous biochar is a promising carbon material for supercapacitor electrodes due to its developed pore structure and high chemical activity which greatly affect the capacitive performance. Predicting the capacitance and exploring the most influential factors are of great significance because it can not only avoid the trial-and-error experiments but also provide guidance for the synthesis of biochar with the aim of capacitance enhancement. In this study, a CNN model with ReLU activation function was established using DenseNet architecture for specific capacitance prediction. The importance and impacts of the physiochemical properties of N-doped porous biochar to the capacitance were revealed. With the guidance of the model, N-doped porous biochar samples with high capacitance were synthesized, the data of which were further used for model validation. This study provides not only a deep learning model which can be used in practice for capacitance prediction but also directions for the synthesis of N-doped porous biochar with high capacitive performance.

Identifying First-Trimester Risk Factors for SGA-LGA Using Weighted Inheritance Voting Ensemble Learning

The classification of fetuses as Small for Gestational Age (SGA) and Large for Gestational Age (LGA) is a critical aspect of neonatal health assessment. SGA and LGA, terms used to describe fetal weights that fall below or above the expected weights for Appropriate for Gestational Age (AGA) fetuses, indicate intrauterine growth restriction and excessive fetal growth, respectively. Early prediction and assessment of latent risk factors associated with these classifications can facilitate timely medical interventions, thereby optimizing the health outcomes for both the infant and the mother. This study aims to leverage first-trimester data to achieve these objectives. This study analyzed data from 7943 pregnant women, including 424 SGA, 928 LGA, and 6591 AGA cases, collected from 2015 to 2021 at the Third Affiliated Hospital of Sun Yat-sen University in Guangzhou, China. We propose a novel algorithm, named the Weighted Inheritance Voting Ensemble Learning Algorithm (WIVELA), to predict the classification of fetuses into SGA, LGA, and AGA categories based on biochemical parameters, maternal factors, and morbidity during pregnancy. Additionally, we proposed algorithms for relevance determination based on the classifier to ascertain the importance of features associated with SGA and LGA. The proposed classification solution demonstrated a notable average accuracy rate of 92.12% on 10-fold cross-validation over 100 loops, outperforming five state-of-the-art machine learning algorithms. Furthermore, we identified significant latent maternal risk factors directly associated with SGA and LGA conditions, such as weight change during the first trimester, prepregnancy weight, height, age, and obstetric factors like fetal growth restriction and birthing LGA baby. This study also underscored the importance of biomarker features at the end of the first trimester, including HDL, TG, OGTT-1h, OGTT-0h, OGTT-2h, TC, FPG, and LDL, which reflect the status of SGA or LGA fetuses. This study presents innovative solutions for classifying and identifying relevant attributes, offering valuable tools for medical teams in the clinical monitoring of fetuses predisposed to SGA and LGA conditions during the initial stage of pregnancy. These proposed solutions facilitate early intervention in nutritional care and prenatal healthcare, thereby contributing to enhanced strategies for managing the health and well-being of both the fetus and the expectant mother.

Identification of subgroups and development of prognostic risk models along the glycolysis-cholesterol synthesis axis in lung adenocarcinoma

Lung cancer is one of the most dangerous malignant tumors affecting human health. Lung adenocarcinoma (LUAD) is the most common subtype of lung cancer. Both glycolytic and cholesterogenic pathways play critical roles in metabolic adaptation to cancer. A dataset of 585 LUAD samples was downloaded from The Cancer Genome Atlas database. We obtained co-expressed glycolysis and cholesterogenesis genes by selecting and clustering genes from Molecular Signatures Database v7.5. We compared the prognosis of different subtypes and identified differentially expressed genes between subtypes. Predictive outcome events were modeled using machine learning, and the top 9 most important prognostic genes were selected by Shapley additive explanation analysis. A risk score model was built based on multivariate Cox analysis. LUAD patients were categorized into four metabolic subgroups: cholesterogenic, glycolytic, quiescent, and mixed. The worst prognosis was the mixed subtype. The prognostic model had great predictive performance in the test set. Patients with LUAD were effectively typed by glycolytic and cholesterogenic genes and were identified as having the worst prognosis in the glycolytic and cholesterogenic enriched gene groups. The prognostic model can provide an essential basis for clinicians to predict clinical outcomes for patients. The model was robust on the training and test datasets and had a great predictive performance.

Evaluating the prognostic value of tumor deposits in non-metastatic lymph node-positive colon adenocarcinoma using Cox regression and machine learning

BACKGROUND

The 8th AJCC TNM staging for non-metastatic lymph node-positive colon adenocarcinoma patients(NMLP-CA) stages solely by lymph node status, irrespective of the positivity of tumor deposits (TD). This study uses machine learning and Cox regression to predict the prognostic value of tumor deposits in NMLP-CA.

METHODS

Patient data from the SEER registry (2010-2019) was used to develop CSS nomograms based on prognostic factors identified via multivariate Cox regression. Model performance was evaluated by c-index, dynamic calibration, and Schmid score. Shapley additive explanations (SHAP) were used to explain the selected models.

RESULTS

The study included 16,548 NMLP-CA patients, randomized 7:3 into training (n = 11,584) and test (n = 4964) sets. Multivariate Cox analysis identified TD, age, marital status, primary site, grade, pT stage, and pN stage as prognostic for cancer-specific survival (CSS). In the test set, the gradient boosting machine (GBM) model achieved the best C-index (0.733) for CSS prediction, while the Cox model and GAMBoost model optimized dynamic calibration(6.473) and Schmid score (0.285), respectively. TD ranked among the top 3 most important features in the models, with increasing predictive significance over time.

CONCLUSIONS

Positive tumor deposit status confers worse prognosis in NMLP-CA patients. Tumor deposits may confer higher TNM staging. Furthermore, TD could play a more significant role in the staging system.

A machine learning model predicts stroke associated with blood cadmium level

Stroke is the leading cause of death and disability worldwide. Cadmium is a prevalent environmental toxicant that may contribute to cardiovascular disease, including stroke. We aimed to build an effective and interpretable machine learning (ML) model that links blood cadmium to the identification of stroke. Our data exploring the association between blood cadmium and stroke came from the National Health and Nutrition Examination Survey (NHANES, 2013-2014). In total, 2664 participants were eligible for this study. We divided these data into a training set (80%) and a test set (20%). To analyze the relationship between blood cadmium and stroke, a multivariate logistic regression analysis was performed. We constructed and tested five ML algorithms including K-nearest neighbor (KNN), decision tree (DT), logistic regression (LR), multilayer perceptron (MLP), and random forest (RF). The best-performing model was selected to identify stroke in US adults. Finally, the features were interpreted using the Shapley Additive exPlanations (SHAP) tool. In the total population, participants in the second, third, and fourth quartiles had an odds ratio of 1.32 (95% CI 0.55, 3.14), 1.65 (95% CI 0.71, 3.83), and 2.67 (95% CI 1.10, 6.49) for stroke compared with the lowest reference group for blood cadmium, respectively. This blood cadmium-based LR approach demonstrated the greatest performance in identifying stroke (area under the operator curve: 0.800, accuracy: 0.966). Employing interpretable methods, we found blood cadmium to be a notable contributor to the predictive model. We found that blood cadmium was positively correlated with stroke risk and that stroke risk from cadmium exposure could be effectively predicted by using ML modeling.

Personalized prediction of mortality in patients with acute ischemic stroke using explainable artificial intelligence

BACKGROUND

Research into the acute kidney disease (AKD) after acute ischemic stroke (AIS) is rare, and how clinical features influence its prognosis remain unknown. We aim to employ interpretable machine learning (ML) models to study AIS and clarify its decision-making process in identifying the risk of mortality.

METHODS

We conducted a retrospective cohort study involving AIS patients from January 2020 to June 2021. Patient data were randomly divided into training and test sets. Eight ML algorithms were employed to construct predictive models for mortality. The performance of the best model was evaluated using various metrics. Furthermore, we created an artificial intelligence (AI)-driven web application that leveraged the top ten most crucial features for mortality prediction.

RESULTS

The study cohort consisted of 1633 AIS patients, among whom 257 (15.74%) developed subacute AKD, 173 (10.59%) experienced AKI recovery, and 65 (3.98%) met criteria for both AKI and AKD. The mortality rate stood at 4.84%. The LightGBM model displayed superior performance, boasting an AUROC of 0.96 for mortality prediction. The top five features linked to mortality were ACEI/ARE, renal function trajectories, neutrophil count, diuretics, and serum creatinine. Moreover, we designed a web application using the LightGBM model to estimate mortality risk.

CONCLUSIONS

Complete renal function trajectories, including AKI and AKD, are vital for fitting mortality in AIS patients. An interpretable ML model effectively clarified its decision-making process for identifying AIS patients at risk of mortality. The AI-driven web application has the potential to contribute to the development of personalized early mortality prevention.

Prediction of Adolescents' Fluid Intelligence Scores based on Deep Learning with Reconstruction Regularization

Objective

The aim of this study was to develop a predictive model for uncorrected/actual fluid intelligence scores in 9-10 year old children using magnetic resonance T1-weighted imaging. Explore the predictive performance of an autoencoder model based on reconstruction regularization for fluid intelligence in adolescents.

Methods

We collected actual fluid intelligence scores and T1-weighted MRIs of 11,534 adolescents who completed baseline tasks from ABCD Data Release 3.0. A total of 148 ROIs were selected and 604 features were proposed by FreeSurfer segmentation. The training and testing sets were divided in a ratio of 7:3. To predict fluid intelligence scores, we used AE, MLP and classic machine learning models, and compared their performance on the test set. In addition, we explored their performance across gender subpopulations. Moreover, we evaluated the importance of features using the SHapley Additive Explain method. Results: The proposed model achieves optimal performance on the test set for predicting actual fluid intelligence scores (PCC = 0.209 ± 0.02, MSE = 105.212 ± 2.53). Results show that autoencoders with refactoring regularization are significantly more effective than MLPs and classical machine learning models. In addition, all models performed better on female adolescents than on male adolescents. Further analysis of relevant characteristics in different populations revealed that this may be related to gender differences in underlying fluid intelligence mechanisms.

Conclusions

We construct a weak but stable correlation between brain structural features and raw fluid intelligence using autoencoders. Future research may need to explore ensemble regression strategies utilizing multiple machine learning algorithms on multimodal data in order to improve the predictive performance of fluid intelligence based on neuroimaging features.

Interpretability-based machine learning for predicting the risk of death from pulmonary inflammation in Chinese intensive care unit patients

Objective

The objective of this research was to create a machine learning predictive model that could be easily interpreted in order to precisely determine the risk of premature death in patients receiving intensive care after pulmonary inflammation.

Methods

In this study, information from the China intensive care units (ICU) Open Source database was used to examine data from 2790 patients who had infections between January 2019 and December 2020. A 7:3 ratio was used to randomly assign the whole patient population to training and validation groups. This study used six machine learning techniques: logistic regression, random forest, gradient boosting tree, extreme gradient boosting tree (XGBoost), multilayer perceptron, and K-nearest neighbor. A cross-validation grid search method was used to search the parameters in each model. Eight metrics were used to assess the models' performance: accuracy, precision, recall, F1 score, area under the curve (AUC) value, Brier score, Jordon's index, and calibration slope. The machine methods were ranked based on how well they performed in each of these metrics. The best-performing models were selected for interpretation using both the Shapley Additive exPlanations (SHAP) and Local interpretable model-agnostic explanations (LIME) interpretable techniques.

Results

A subset of the study cohort's patients (120/1668, or 7.19%) died in the hospital following screening for inclusion and exclusion criteria. Using a cross-validated grid search to evaluate the six machine learning techniques, XGBoost showed good discriminative ability, achieving an accuracy score of 0.889 (0.874-0.904), precision score of 0.871 (0.849-0.893), recall score of 0.913 (0.890-0.936), F1 score of 0.891 (0.876-0.906), and AUC of 0.956 (0.939-0.973). Additionally, XGBoost exhibited excellent performance with a Brier score of 0.050, Jordon index of 0.947, and calibration slope of 1.074. It was also possible to create an interactive internet page using the XGBoost model.

Conclusion

By identifying patients at higher risk of early mortality, machine learning-based mortality risk prediction models have the potential to significantly improve patient care by directing clinical decision making and enabling early detection of survival and mortality issues in patients with pulmonary inflammation disease.

Data-Driven Approaches to Predict Dendrimer Cytotoxicity

Dendrimers are employed as functional elements in contrast agents and are proposed as nontoxic vehicles for drug delivery. Toxicity is a property that is to be evaluated for this novel class of bionanomaterials for in vivo applications. The current research is hampered due to the lack of structured data sets for toxicity studies for dendrimers. In this work, we have built a data set by curating literature for toxicity data and augmented it with structural and physicochemical features. We present a comprehensive, feature-rich database of dendrimer toxicity measured across various cell lines for prediction, design, and optimization studies. We have also explored novel computational approaches for predicting dendrimer cytotoxicity. We demonstrate superior outcomes for toxicity prediction using essential regression in the space of small data sets.

Implementation of a Machine Learning Approach Evaluating Risk Factors for Complications after Single-Stage Augmentation Mastopexy

BACKGROUND

Single-stage mastopexy augmentation is a much-debated intervention due to its complexity and the associated relatively high complication rates. This study aimed to reevaluate the risk factors for these complications using a novel approach based on artificial intelligence and to demonstrate its possible limitations.

PATIENTS AND METHODS

Complete datasets of patients who underwent single-staged augmentation mastopexy during 2014-2023 at one institution by a single surgeon were collected retrospectively. These were subsequently processed and analyzed by CART, RF and XGBoost algorithms.

RESULTS

A total of 342 patients were included in the study, of which 43 (12.57%) reported surgery-associated complications, whereby capsular contracture (n = 19) was the most common. BMI represented the most important variable for the development of complications (FIS = 0.44 in CART). 2.9% of the patients expressed the desire for implant change in the course, with absence of any complications. A statistically significant correlation between smoking and the desire for implant change (p < 0.001) was revealed.

CONCLUSION

The importance of implementing artificial intelligence into clinical research could be underpinned by this study, as risk variables can be reclassified based on factors previously considered less or even irrelevant. Thereby we encountered limitations using ML approaches. Further studies will be needed to investigate the association between smoking, BMI and the current implant size with the desire for implant change without any complications. Moreover, we could show that the procedure can be performed safely without high risk of developing major complications.

LEVEL OF EVIDENCE IV

This journal requires that authors assign a level of evidence to each article. For a full description of these Evidence-Based Medicine ratings, please refer to the Table of Contents or the online Instructions to Authors www.springer.com/00266.

Novel machine learning insights into the QM7b and QM9 quantum mechanics datasets

This paper (i) explores the internal structure of two quantum mechanics datasets (QM7b, QM9), composed of several thousands of organic molecules and described in terms of electronic properties, and (ii) further explores an inverse design approach to molecular design consisting of using machine learning methods to approximate the atomic composition of molecules, using QM9 data. Understanding the structure and characteristics of this kind of data is important when predicting the atomic composition from physical-chemical properties in inverse molecular designs. Intrinsic dimension analysis, clustering, and outlier detection methods were used in the study. They revealed that for both datasets the intrinsic dimensionality is several times smaller than the descriptive dimensions. The QM7b data is composed of well-defined clusters related to atomic composition. The QM9 data consists of an outer region predominantly composed of outliers, and an inner, core region that concentrates clustered inliner objects. A significant relationship exists between the number of atoms in the molecule and its outlier/inliner nature. The spatial structure exhibits a relationship with molecular weight. Despite the structural differences between the two datasets, the predictability of variables of interest for inverse molecular design is high. This is exemplified by models estimating the number of atoms of the molecule from both the original properties and from lower dimensional embedding spaces. In the generative approach the input is given by a set of desired properties of the molecule and the output is an approximation of the atomic composition in terms of its constituent chemical elements. This could serve as the starting region for further search in the huge space determined by the set of possible chemical compounds. The quantum mechanic's dataset QM9 is used in the study, composed of 133,885 small organic molecules and 19 electronic properties. Different multi-target regression approaches were considered for predicting the atomic composition from the properties, including feature engineering techniques in an auto-machine learning framework. High-quality models were found that predict the atomic composition of the molecules from their electronic properties, as well as from a subset of only 52.6% size. Feature selection worked better than feature generation. The results validate the generative approach to inverse molecular design.

Forecasting Myopic Maculopathy Risk Over a Decade: Development and Validation of an Interpretable Machine Learning Algorithm

Purpose

The purpose of this study was to develop and validate prediction model for myopic macular degeneration (MMD) progression in patients with high myopia.

Methods

The Zhongshan High Myopia Cohort for model development included 660 patients aged 7 to 70 years with a bilateral sphere of ≤-6.00 diopters (D). Two hundred twelve participants with an axial length (AL) ≥25.5 mm from the Chinese Ocular Imaging Project were used for external validation. Thirty-four clinical variables, including demographics, lifestyle, myopia history, and swept source optical coherence tomography data, were analyzed. Sequential forward selection was used for predictor selection, and binary classification models were created using five machine learning algorithms to forecast the risk of MMD progression over 10 years.

Results

Over a median follow-up of 10.9 years, 133 patients (20.2%) showed MMD progression in the development cohort. Among them, 69 (51.9%) developed newly-onset MMD, 11 (8.3%) developed patchy atrophy from diffuse atrophy, 54 (40.6%) showed an enlargement of lesions, and 9 (6.8%) developed plus signs. Top six predictors for MMD progression included thinner subfoveal choroidal thickness, longer AL, worse best-corrected visual acuity, older age, female gender, and shallower anterior chamber depth. The eXtreme Gradient Boosting algorithm yielded the best discriminative performance (area under the receiver operating characteristic curve [AUROC] = 0.87 ± 0.02) with good calibration in the training cohort. In a less myopic external validation group (median -5.38 D), 48 patients (22.6%) developed MMD progression over 4 years, with the model's AUROC validated at 0.80 ± 0.008.

Conclusions

Machine learning model effectively predicts MMD progression a decade ahead using clinical and imaging indicators. This tool shows promise for identifying "at-risk" high myopes for timely intervention and vision protection.

Artificial intelligence to unlock real-world evidence in clinical oncology: A primer on recent advances

PURPOSE

Real world evidence is crucial to understanding the diffusion of new oncologic therapies, monitoring cancer outcomes, and detecting unexpected toxicities. In practice, real world evidence is challenging to collect rapidly and comprehensively, often requiring expensive and time-consuming manual case-finding and annotation of clinical text. In this Review, we summarise recent developments in the use of artificial intelligence to collect and analyze real world evidence in oncology.

METHODS

We performed a narrative review of the major current trends and recent literature in artificial intelligence applications in oncology.

RESULTS

Artificial intelligence (AI) approaches are increasingly used to efficiently phenotype patients and tumors at large scale. These tools also may provide novel biological insights and improve risk prediction through multimodal integration of radiographic, pathological, and genomic datasets. Custom language processing pipelines and large language models hold great promise for clinical prediction and phenotyping.

CONCLUSIONS

Despite rapid advances, continued progress in computation, generalizability, interpretability, and reliability as well as prospective validation are needed to integrate AI approaches into routine clinical care and real-time monitoring of novel therapies.

Multicentre validation of a machine learning model for predicting respiratory failure after noncardiac surgery

BACKGROUND

Postoperative respiratory failure is a serious complication that could benefit from early accurate identification of high-risk patients. We developed and validated a machine learning model to predict postoperative respiratory failure, defined as prolonged (>48 h) mechanical ventilation or reintubation after surgery.

METHODS

Easily extractable electronic health record (EHR) variables that do not require subjective assessment by clinicians were used. From EHR data of 307,333 noncardiac surgical cases, the model, trained with a gradient boosting algorithm, utilised a derivation cohort of 99,025 cases from Seoul National University Hospital (2013-9). External validation was performed using three separate cohorts A-C from different hospitals comprising 208,308 cases. Model performance was assessed by area under the receiver operating characteristic (AUROC) curve and area under the precision-recall curve (AUPRC), a measure of sensitivity and precision at different thresholds.

RESULTS

The model included eight variables: serum albumin, age, duration of anaesthesia, serum glucose, prothrombin time, serum creatinine, white blood cell count, and body mass index. Internally, the model achieved an AUROC of 0.912 (95% confidence interval [CI], 0.908-0.915) and AUPRC of 0.113. In external validation cohorts A, B, and C, the model achieved AUROCs of 0.879 (95% CI, 0.876-0.882), 0.872 (95% CI, 0.870-0.874), and 0.931 (95% CI, 0.925-0.936), and AUPRCs of 0.029, 0.083, and 0.124, respectively.

CONCLUSIONS

Utilising just eight easily extractable variables, this machine learning model demonstrated excellent discrimination in both internal and external validation for predicting postoperative respiratory failure. The model enables personalised risk stratification and facilitates data-driven clinical decision-making.

A precise blood transfusion evaluation model for aortic surgery: a single-center retrospective study

Cardiac aortic surgery is an extremely complicated procedure that often requires large volume blood transfusions during the operation. Currently, it is not possible to accurately estimate the intraoperative blood transfusion volume before surgery. Therefore, in this study, to determine the clinically precise usage of blood for intraoperative blood transfusions during aortic surgery, we established a predictive model based on machine learning algorithms. We performed a retrospective analysis on 4,285 patients who received aortic surgery in Beijing Anzhen Hospital between January 2018 and September 2022. Ultimately, 3,654 patients were included in the study, including 2,557 in the training set and 1,097 in the testing set. By utilizing 13 current mainstream models and a large-scale cardiac aortic surgery dataset, we built a novel machine learning model for accurately predicting intraoperative red blood cell transfusion volume. Based on the transfusion-related risk factors that the model identified, we also established the relevant variables that affected the results. The results revealed that decision tree models were the most suitable for predicting the blood transfusion volume during aortic surgery. In particular, the mean absolute error for the best-performing extremely randomized forest model was 1.17 U, while the R2 value was 0.50. Further exploration into intraoperative blood transfusion during aortic surgery identified erythrocytes, estimated operation duration, body weight, sex, red blood cell count, and D-dimer as the six most significant risk factors. These factors were subsequently analyzed for their influence on intraoperative blood transfusion volume in relevant patients, as well as the protective threshold for prediction. The novel intraoperative blood transfusion prediction model for cardiac aorta surgery in this study effectively assists clinicians in accurately calculating blood transfusion volumes and achieving effective utilization of blood resources. Furthermore, we utilize interpretability technology to reveal the influence of critical risk factors on intraoperative blood transfusion volume, which provides an important reference for physicians to provide timely and effective interventions. It also enables personalized and precise intraoperative blood usage.

Interpretable machine learning model for predicting acute kidney injury in critically ill patients

BACKGROUND

This study aimed to create a method for promptly predicting acute kidney injury (AKI) in intensive care patients by applying interpretable, explainable artificial intelligence techniques.

METHODS

Population data regarding intensive care patients were derived from the Medical Information Mart for Intensive Care IV database from 2008 to 2019. Machine learning (ML) techniques with six methods were created to construct the predicted models for AKI. The performance of each ML model was evaluated by comparing the areas under the curve (AUC). Local Interpretable Model-Agnostic Explanations (LIME) method and Shapley Additive exPlanation values were used to decipher the best model.

RESULTS

According to inclusion and exclusion criteria, 53,150 severely sick individuals were included in the present study, of which 42,520 (80%) were assigned to the training group, and 10,630 (20%) were allocated to the validation group. Compared to the other five ML models, the eXtreme Gradient Boosting (XGBoost) model greatly predicted AKI following ICU admission, with an AUC of 0.816. The top four contributing variables of the XGBoost model were SOFA score, weight, mechanical ventilation, and the Simplified Acute Physiology Score II. An AKI and Non-AKI cases were predicted separately using the LIME algorithm.

CONCLUSION

Overall, the constructed clinical feature-based ML models are excellent in predicting AKI in intensive care patients. It would be constructive for physicians to provide early support and timely intervention measures to intensive care patients at risk of AKI.