Evaluation of a decided sample size in machine learning applications

Advancing the diagnosis of major depressive disorder: Integrating neuroimaging and machine learning

Major depressive disorder (MDD), a psychiatric disorder characterized by functional brain deficits, poses considerable diagnostic and treatment challenges, especially in adolescents owing to varying clinical presentations. Biomarkers hold substantial clinical potential in the field of mental health, enabling objective assessments of physiological and pathological states, facilitating early diagnosis, and enhancing clinical decision-making and patient outcomes. Recent breakthroughs combine neuroimaging with machine learning (ML) to distinguish brain activity patterns between MDD patients and healthy controls, paving the way for diagnostic support and personalized treatment. However, the accuracy of the results depends on the selection of neuroimaging features and algorithms. Ensuring privacy protection, ML model accuracy, and fostering trust are essential steps prior to clinical implementation. Future research should prioritize the establishment of comprehensive legal frameworks and regulatory mechanisms for using ML in MDD diagnosis while safeguarding patient privacy and rights. By doing so, we can advance accuracy and personalized care for MDD.

A cross-dataset study on automatic detection of autism spectrum disorder from text data

OBJECTIVE

The goals of this article are as follows. First, to investigate the possibility of detecting autism spectrum disorder (ASD) from text data using the latest generation of machine learning tools. Second, to compare model performance on two datasets of transcribed statements, collected using two different diagnostic tools. Third, to investigate the feasibility of knowledge transfer between models trained on both datasets and check if data augmentation can help alleviate the problem of a small number of observations.

METHOD

We explore two techniques to detect ASD. The first one is based on fine-tuning HerBERT, a BERT-based, monolingual deep transformer neural network. The second one uses the newest, multipurpose text embeddings from OpenAI and a classifier. We apply the methods to two separate datasets of transcribed statements, collected using two different diagnostic tools: thought, language, and communication (TLC) and autism diagnosis observation schedule-2 (ADOS-2). We conducted several cross-dataset experiments in both a zero-shot setting and a setting where models are pretrained on one dataset and then training continues on another to test the possibility of knowledge transfer.

RESULTS

Unlike previous studies, the models we tested obtained average results on ADOS-2 data but reached very good performance of the models in TLC. We did not observe any benefits from knowledge transfer between datasets. We observed relatively poor performance of models trained on augmented data and hypothesize that data augmentation by back translation obfuscates autism-specific signals.

CONCLUSION

The quality of machine learning models that detect ASD from text data is improving, but model results are dependent on the type of input data or diagnostic tool.

Proteomic associations with cognitive variability as measured by the Wisconsin Card Sorting Test in a healthy Thai population: A machine learning approach

Inter-individual cognitive variability, influenced by genetic and environmental factors, is crucial for understanding typical cognition and identifying early cognitive disorders. This study investigated the association between serum protein expression profiles and cognitive variability in a healthy Thai population using machine learning algorithms. We included 199 subjects, aged 20 to 70, and measured cognitive performance with the Wisconsin Card Sorting Test. Differentially expressed proteins (DEPs) were identified using label-free proteomics and analyzed with the Linear Model for Microarray Data. We discovered 213 DEPs between lower and higher cognition groups, with 155 upregulated in the lower cognition group and enriched in the IL-17 signaling pathway. Subsequent bioinformatic analysis linked these DEPs to neuroinflammation-related cognitive impairment. A random forest model classified cognitive ability groups with an accuracy of 81.5%, sensitivity of 65%, specificity of 85.9%, and an AUC of 0.79. By targeting a specific Thai cohort, this research provides novel insights into the link between neuroinflammation and cognitive performance, advancing our understanding of cognitive variability, highlighting the role of biological markers in cognitive function, and contributing to developing more accurate machine learning models for diverse populations.

Machine learning classification of active viewing of pain and non-pain images using EEG does not exceed chance in external validation samples

Previous research has demonstrated that machine learning (ML) could not effectively decode passive observation of neutral versus pain photographs by using electroencephalogram (EEG) data. Consequently, the present study explored whether active viewing, i.e., requiring participant engagement in a task, of neutral and pain stimuli improves ML performance. Random forest (RF) models were trained on cortical event-related potentials (ERPs) during a two-alternative forced choice paradigm, whereby participants determined the presence or absence of pain in photographs of facial expressions and action scenes. Sixty-two participants were recruited for the model development sample. Moreover, a within-subject temporal validation sample was collected, consisting of 27 subjects. In line with our previous research, three RF models were developed to classify images into faces and scenes, neutral and pain scenes, and neutral and pain expressions. The results demonstrated that the RF successfully classified discrete categories of visual stimuli (faces and scenes) with accuracies of 78% and 66% on cross-validation and external validation, respectively. However, despite promising cross-validation results of 61% and 67% for the classification of neutral and pain scenes and neutral and pain faces, respectively, the RF models failed to exceed chance performance on the external validation dataset on both empathy classification attempts. These results align with previous research, highlighting the challenges of classifying complex states, such as pain empathy using ERPs. Moreover, the results suggest that active observation fails to enhance ML performance beyond previous passive studies. Future research should prioritise improving model performance to obtain levels exceeding chance, which would demonstrate increased utility.

Estimating the Minimum Sample Size for Neural Network Model Fitting-A Monte Carlo Simulation Study

In the era of machine learning, many psychological studies use machine learning methods. Specifically, neural networks, a set of machine learning methods that exhibit exceptional performance in various tasks, have been used on psychometric datasets for supervised model fitting. From the computer scientist's perspective, psychometric independent variables are typically ordinal and low-dimensional-characteristics that can significantly impact model performance. To our knowledge, there is no guidance about the sample planning suggestion for this task. Therefore, we conducted a simulation study to test the performance of an NN with different sample sizes and the simulation of both linear and nonlinear relationships. We proposed the minimum sample size for the neural network model fitting with two criteria: the performance of 95% of the models is close to the theoretical maximum, and 80% of the models can outperform the linear model. The findings of this simulation study show that the performance of neural networks can be unstable with ordinal variables as independent variables, and we suggested that neural networks should not be used on ordinal independent variables with at least common nonlinear relationships in psychology. Further suggestions and research directions are also provided.

Do Machine Learning Approaches Perform Better Than Regression Models in Mapping Studies? A Systematic Review

OBJECTIVES

To identify how machine learning (ML) approaches were implemented in mapping studies and to determine the extent to which ML improved performance compared with regression models (RMs).

METHODS

A systematic literature search was conducted in 12 databases from inception to December 2023 to identify studies that applied ML to develop mapping algorithms. A data template was applied to extract data set information, source and target measures, ML approaches and RMs, mapping types (direct vs indirect), goodness-of-fit indicators (mean absolute error, mean squared error, root mean squared error, R-squared, and intraclass correlation coefficient), and validation methods. Differences in goodness-of-fit indicators between ML and RMs were summarized. Potential advantages and challenges for ML were further discussed.

RESULTS

Thirteen mapping studies were identified, in which both ML and RM were adopted. Bayesian networks were the most frequently used ML approach (n = 6), followed by the least absolute shrinkage and selection operator (n = 4). The ordinary least square model was the most used RM (n = 8), followed by the censored least absolute deviation and multinomial logit models (n = 5 each). The average improvement in the goodness-of-fit of ML compared with that of RMs by indicators were 0.007 (mean absolute error), 0.004 (mean squared error), 0.058 (R-squared), 0.016 (intraclass correlation coefficient), and -0.0004 (root mean squared error).

CONCLUSIONS

There is an increasing number of studies using ML in developing mapping algorithms. Generally, a minor improvement of goodness-of-fit was observed compared with RMs when using mean-based comparisons. Issues such as how to interpret, apply, and externally validate the ML-based outputs would affect their implementation. Future studies are warranted to verify advantages of ML approaches.

Differentiating between renal medullary and clear cell renal carcinoma with a machine learning radiomics approach

BACKGROUND

The objective of this study was to develop and validate a radiomics-based machine learning (ML) model to differentiate between renal medullary carcinoma (RMC) and clear cell renal carcinoma (ccRCC).

METHODS

This retrospective Institutional Review Board -approved study analyzed CT images and clinical data from patients with RMC (n = 87) and ccRCC (n = 93). Patients without contrast-enhanced CT scans obtained before nephrectomy were excluded. A standard volumetric software package (MIM 7.1.4, MIM Software Inc.) was used for contouring, after which 949 radiomics features were extracted with PyRadiomics 3.1.0. Radiomics analysis was then performed with RadAR for differential radiomics analysis. ML was then performed with extreme gradient boosting (XGBoost 2.0.3) to differentiate between RMC and ccRCC. Three separate ML models were created to differentiate between ccRCC and RMC. These models were based on clinical demographics, radiomics, and radiomics incorporating hemoglobin electrophoresis for sickle cell trait, respectively.

RESULTS

Performance metrics for the 3 developed ML models were as follows: demographic factors only (AUC = 0.777), calibrated radiomics (AUC = 0.915), and calibrated radiomics with sickle cell trait incorporated (AUC = 1.0). The top 4 ranked features from differential radiomic analysis, ranked by their importance, were run entropy (preprocessing filter = original, AUC = 0.67), dependence entropy (preprocessing filter = wavelet, AUC = 0.67), zone entropy (preprocessing filter = original, AUC = 0.67), and dependence entropy (preprocessing filter = original, AUC = 0.66).

CONCLUSION

A radiomics-based machine learning model effectively differentiates between ccRCC and RMC. This tool can facilitate the radiologist's ability to suspicion and decrease the misdiagnosis rate of RMC.

Machine learning research methods to predict postoperative pain and opioid use: a narrative review

The use of machine learning to predict postoperative pain and opioid use has likely been catalyzed by the availability of complex patient-level data, computational and statistical advancements, the prevalence and impact of chronic postsurgical pain, and the persistence of the opioid crisis. The objectives of this narrative review were to identify and characterize methodological aspects of studies that have developed and/or tested machine learning algorithms to predict acute, subacute, or chronic pain or opioid use after any surgery and to propose considerations for future machine learning studies. Pairs of independent reviewers screened titles and abstracts of 280 PubMed-indexed articles and ultimately extracted data from 61 studies that met entry criteria. We observed a marked increase in the number of relevant publications over time. Studies most commonly focused on machine learning algorithms to predict chronic postsurgical pain or opioid use, using real-world data from patients undergoing orthopedic surgery. We identified variability in sample size, number and type of predictors, and how outcome variables were defined. Patient-reported predictors were highlighted as particularly informative and important to include in such machine learning algorithms, where possible. We hope that findings from this review might inform future applications of machine learning that improve the performance and clinical utility of resultant machine learning algorithms.

The diagnostic and prognostic capability of artificial intelligence in spinal cord injury: A systematic review

Background

Artificial intelligence (AI) models have shown potential for diagnosing and prognosticating traumatic spinal cord injury (tSCI), but their clinical utility remains uncertain.

Method

ology: The primary aim was to evaluate the performance of AI algorithms in diagnosing and prognosticating tSCI. Subsequent systematic searching of seven databases identified studies evaluating AI models. PROBAST and TRIPOD tools were used to assess the quality and reporting of included studies (PROSPERO: CRD42023464722). Fourteen studies, comprising 20 models and 280,817 pooled imaging datasets, were included. Analysis was conducted in line with the SWiM guidelines.

Results

For prognostication, 11 studies predicted outcomes including AIS improvement (30%), mortality and ambulatory ability (20% each), and discharge or length of stay (10%). The mean AUC was 0.770 (range: 0.682-0.902), indicating moderate predictive performance. Diagnostic models utilising DTI, CT, and T2-weighted MRI with CNN-based segmentation achieved a weighted mean accuracy of 0.898 (range: 0.813-0.938), outperforming prognostic models.

Conclusion

AI demonstrates strong diagnostic accuracy (mean accuracy: 0.898) and moderate prognostic capability (mean AUC: 0.770) for tSCI. However, the lack of standardised frameworks and external validation limits clinical applicability. Future models should integrate multimodal data, including imaging, patient characteristics, and clinician judgment, to improve utility and alignment with clinical practice.

Understanding Daily Care Experience Preferences Across the Lifespan of Older Adults: Application of Natural Language Processing

INTRODUCTION

Older adults are a heterogeneous group, and their care experience preferences are likely to be diverse and individualized. Thus, the aim of this study was to identify categories of older adults' care experience preferences and to examine similarities and differences across different age groups.

METHODS

The initial categories of older adults' care experience preferences were identified through a qualitative review of narrative text (n = 3134) in the ADVault data set. A natural language processing (NLP) algorithm was used to automatically and accurately define older adults' care experience preference categories. Descriptive statistics were used to examine similarities and differences in care experience preference categories across different age groups.

RESULTS

The overall average performance of NLP algorithms was relatively high (average F-score = 0.88; range: 0.77-0.96). Through a qualitative review of 350 randomly selected texts, a total of 11 categories were identified. The most frequent category was music, followed by photographs, entertainment, family/friends, religion-related, atmosphere, flower/plants, pet, bed/bedding, hobby, and other. After applying the best performing NLP algorithm to each category, older adults' care experience preference categories were music (41.32%), followed by photographs (28.47%), entertainment (13.46%), religion-related (n = 12.69%), pet (12.22%), flower/plants (11.51%), family/friends (8.45%), atmosphere (7.75%), bed/bedding (6.12%), and hobby (5.45%). Young-old and old-old adults had similar care experience preferences with music being the most frequent category while old-old adults had photographs as the most frequent category for their care experience preference.

CONCLUSION

Clinicians must understand the distinct categories of care experience preferences and incorporate them into personalized care planning.

Multimodal Artificial Intelligence Model for Prediction of Abdominal Aortic Aneurysm Shrinkage After Endovascular Repair ( the ART in EVAR study)

PURPOSE

The goal of the study described in this protocol is to build a multimodal artificial intelligence (AI) model to predict abdominal aortic aneurysm (AAA) shrinkage 1 year after endovascular aneurysm repair (EVAR).

METHODS

In this retrospective observational multicenter study, approximately 1000 patients will be enrolled from hospital records of 5 experienced vascular centers. Patients will be included if they underwent elective EVAR for infrarenal AAA with initial assisted technical success and had imaging available of the same modality preoperatively and at 1-year follow-up (CTA-CTA or US-US). Data collection will include baseline and vascular characteristics, medication use, procedural data, preoperative and postoperative imaging data, follow-up data, and complications.

PROPOSED ANALYSES

The cohort will be stratified into 3 groups of AAA remodeling based on the maximum AAA diameter difference between the preoperative and 1-year postoperative moment. Patients with a diameter reduction of ≥5 mm will be assigned to the AAA shrinkage group, cases with an increase of ≥5 mm will be assigned to the AAA growth group, and patients with a diameter increase or reduction of <5 mm will be assigned to the stable AAA group. Furthermore, an additional fourth group will include all patients who underwent an AAA-related reintervention within the first year after EVAR, because both the complication and the reintervention might have influenced the state of AAA remodeling at 1 year. The preoperative and postoperative CTA scans will be used for anatomical AAA analysis and biomechanical assessment through semi-automatic segmentation and finite element analysis. All collected clinical, biomechanical, and imaging data will be used to create an AI prediction model for AAA shrinkage. Explainable AI techniques will be used to identify the most descriptive input features in the model. Predicting factors resulting from the AI model will be compared with conventional univariate and multivariate logistic regression analyses to find the best model for the prediction of AAA shrinkage. The study is registered at www.clinicaltrials.gov under the registration number NCT06250998.

CLINICAL IMPACT

This study aims to develop a robust and high-performance AI model for predicting AAA shrinkage one-year after EVAR, with great potential for optimizing both EVAR treatment and follow-up. The model can identify cases with an initially lower chance of early AAA shrinkage, in whom EVAR-treatment could be tailored by including additional preoperative coil embolization, active sac management and/or postoperative tranexamic acid therapy, which have shown to promote AAA shrinkage rate but are too complex and costly to perform in all patients. The model could aid in stratification of post-EVAR surveillance based on the patient's individual risk and possibly decrease follow-up for the 40-50% of patients who will experience AAA sac shrinkage. Overall, the AI prediction model is expected to improve patient survival and decrease the number of reinterventions after EVAR and associated healthcare costs.

Predicting home delivery and identifying its determinants among women aged 15-49 years in sub-Saharan African countries using a Demographic and Health Surveys 2016-2023: a machine learning algorithm

BACKGROUND

Birth-related mortality is significantly increased by home births without skilled medical assistance during delivery, presenting a major risk to the public's health. The objective of this study is to predict home delivery and identify the determinants using machine learning algorithm in sub-Saharan African.

METHODS

This study used design science approaches. The data set obtained from demographic health survey in sub-Saharan African weighted sample of 299,759 women was included in the stud. Machine learning models such as Random Forest, Decision Tree, K-Nearest Neighbor, Logistic Regression, Extreme Gradient Boosting, AdaBoost, Artificial Neural Network, and Naive Bayes were used. The predictive model was evaluated by area under the curve, accuracy, precision, recall, and F-measure.

RESULTS

The final experimentation results indicated that random forest model performed the best to predict home delivery with accuracy (83%) and, ROC curve (89%). The Shapley additive explanation features an importance plot optimized for random forest model to identifying the most predictors of home delivery. Association rules findings showed that inadequate antenatal care visits, marital status married, no education, mobile phone, television, electricity, poor wealth index, infrequent television viewing, and rural residence were predictor of home delivery.

CONCLUSION

The random forest machine learning model provides greater predictive power for estimating home delivery risk factors. To reduce the prevalence of home delivery, this finding recommends to emphasis on improving antenatal care services, education, and awareness about health facility delivery.

Deep learning in microbiome analysis: a comprehensive review of neural network models

Microbiome research, the study of microbial communities in diverse environments, has seen significant advances due to the integration of deep learning (DL) methods. These computational techniques have become essential for addressing the inherent complexity and high-dimensionality of microbiome data, which consist of different types of omics datasets. Deep learning algorithms have shown remarkable capabilities in pattern recognition, feature extraction, and predictive modeling, enabling researchers to uncover hidden relationships within microbial ecosystems. By automating the detection of functional genes, microbial interactions, and host-microbiome dynamics, DL methods offer unprecedented precision in understanding microbiome composition and its impact on health, disease, and the environment. However, despite their potential, deep learning approaches face significant challenges in microbiome research. Additionally, the biological variability in microbiome datasets requires tailored approaches to ensure robust and generalizable outcomes. As microbiome research continues to generate vast and complex datasets, addressing these challenges will be crucial for advancing microbiological insights and translating them into practical applications with DL. This review provides an overview of different deep learning models in microbiome research, discussing their strengths, practical uses, and implications for future studies. We examine how these models are being applied to solve key problems and highlight potential pathways to overcome current limitations, emphasizing the transformative impact DL could have on the field moving forward.

Impact of Dataset Size on 3D CNN Performance in Intracranial Hemorrhage Classification

Background: This study aimed to evaluate the effect of sample size on the development of a three-dimensional convolutional neural network (3DCNN) model for predicting the binary classification of three types of intracranial hemorrhage (ICH): intraparenchymal, subarachnoid, and subdural (IPH, SAH, SDH, respectively). Methods: During the training, we compiled all images of each brain computed tomography scan into a single 3D image, which was then fed into the model to classify the presence of ICH. We divided the non-hemorrhage quantities into 20, 30, 40, 50, 100, and 150 and the ICH quantities into 20, 30, 40, and 50. Cross-validation was performed to compute the average area under the curve (AUC) over the last five iterations. The AUC and accuracy were used to evaluate the performance of the models. Results: Fifty patients, each with the three ICH types, and 150 non-hemorrhage cases were enrolled. Larger sample sizes achieved stable and acceptable performance in the artificial intelligence (AI) models, whereas training with a limited number of cases posed the risk of falsely high AUC values or accuracy. The overall trends and fluctuations in AUC values were similar between IPH and SDH but different for SAH. The accuracy of the results was relatively consistent among the three ICH types. Conclusions: The 3DCNN technique can be used to develop AI models capable of detecting ICH from limited case numbers. However, a minimal case number must be provided. The performance of AI models varies across different ICH types and is more stable with larger sample sizes.

Machine learning models for neurocognitive outcome prediction in preterm born infants

BACKGROUND

Outcome prediction after preterm birth is important for long-term neonatal care, but has proven notoriously challenging for neurocognitive outcome. This study investigated the potential of machine learning to improve neurocognitive outcome prediction at two and five years of corrected age in preterm infants, using readily available predictors from the neonatal setting.

METHODS

Predictors originating from the antenatal and neonatal period of preterm infants born <30 weeks gestation were used to predict adverse neurocognitive outcome on the Bayley Scale and Wechsler Preschool and Primary Scale of Intelligence. Machine learning models were compared to conventional logistic regression and validated using internal cross-validation.

RESULTS

Best performing models were a random forest (two-year outcome) and a support vector machine (five-year outcome) with an area under the receiver operating characteristic curve (AUC) of 0.682 and 0.695 respectively, reaching high negative predictive values (95% and 91%, respectively). These models performed significantly better than the conventional models.

CONCLUSIONS

The models reached moderate overall predictive performance, yet with promising potential for early identification of children without adverse neurocognitive outcome. Machine learning modestly improved neurocognitive outcome prediction. Future research may harvest the predictive potential of a wider variety of routine (clinical) data, such as vital sign time series.

IMPACT

Early prediction of neurocognitive outcome in preterm infants will enable targeted follow-up and deployment of early (preventative) interventions to improve outcome. Neurocognitive outcome remains notoriously challenging using conventional models, while existing machine learning models depend on advanced MRI-derived predictors with limited potential for implementation into daily clinical practice. This study developed machine learning models for neurocognitive outcome prediction using predictors that are readily available in neonatal settings. Neurocognitive outcome prediction remains challenging due to low AUC and PPV, however, the models demonstrate high NPV, indicating potential for identifying children at low risk for adverse outcome.

Enhancing Treatment Decisions for Advanced Non-Small Cell Lung Cancer with Epidermal Growth Factor Receptor Mutations: A Reinforcement Learning Approach

BACKGROUND

Although higher-generation TKIs are associated with improved progression-free survival in advanced NSCLC patients with EGFR mutations, the optimal selection of TKI treatment remains uncertain. To address this gap, we developed a web application powered by a reinforcement learning (RL) algorithm to assist in guiding initial TKI treatment decisions.

METHODS

Clinical and mutational data from advanced NSCLC patients were retrospectively collected from 14 medical centers. Only patients with complete data and sufficient follow-up were included. Multiple supervised machine learning models were tested, with the Extra Trees Classifier (ETC) identified as the most effective for predicting progression-free survival. Feature importance scores were calculated by the ETC, and features were then integrated into a Deep Q-Network (DQN) RL algorithm. The RL model was designed to select optimal TKI generation and a treatment line for each patient and was embedded into an open-source web application for experimental clinical use.

RESULTS

In total, 318 cases of EGFR-mutant advanced NSCLC were analyzed, with a median patient age of 63. A total of 52.2% of patients were female, and 83.3% had ECOG scores of 0 or 1. The top three most influential features identified were neutrophil-to-lymphocyte ratio (log-transformed), age (log-transformed), and the treatment line of TKI administration, as tested by the ETC algorithm, with an area under curve (AUC) value of 0.73, whereas the DQN RL algorithm achieved a higher AUC value of 0.80, assigning distinct Q-values across four TKI treatment categories. This supports the decision-making process in the web-based 'EGFR Mutant NSCLC Treatment Advisory System', where clinicians can input patient-specific data to receive tailored recommendations.

CONCLUSIONS

The RL-based web application shows promise in assisting TKI treatment selection for EGFR-mutant advanced NSCLC patients, underscoring the potential for reinforcement learning to enhance decision-making in oncology care.

Harnessing the potential of human induced pluripotent stem cells, functional assays and machine learning for neurodevelopmental disorders

Neurodevelopmental disorders (NDDs) affect 4.7% of the global population and are associated with delays in brain development and a spectrum of impairments that can lead to lifelong disability and even mortality. Identification of biomarkers for accurate diagnosis and medications for effective treatment are lacking, in part due to the historical use of preclinical model systems that do not translate well to the clinic for neurological disorders, such as rodents and heterologous cell lines. Human-induced pluripotent stem cells (hiPSCs) are a promising in vitro system for modeling NDDs, providing opportunities to understand mechanisms driving NDDs in human neurons. Functional assays, including patch clamping, multielectrode array, and imaging-based assays, are popular tools employed with hiPSC disease models for disease investigation. Recent progress in machine learning (ML) algorithms also presents unprecedented opportunities to advance the NDD research process. In this review, we compare two-dimensional and three-dimensional hiPSC formats for disease modeling, discuss the applications of functional assays, and offer insights on incorporating ML into hiPSC-based NDD research and drug screening.

Early prediction of long COVID-19 syndrome persistence at 12 months after hospitalisation: a prospective observational study from Ukraine

OBJECTIVE

To identify the early predictors of a self-reported persistence of long COVID syndrome (LCS) at 12 months after hospitalisation and to propose the prognostic model of its development.

DESIGN

A combined cross-sectional and prospective observational study.

SETTING

A tertiary care hospital.

PARTICIPANTS

221 patients hospitalised for COVID-19 who have undergone comprehensive clinical, sonographic and survey-based evaluation predischarge and at 1 month with subsequent 12-month follow-up. The final cohort included 166 patients who had completed the final visit at 12 months.

MAIN OUTCOME MEASURE

A self-reported persistence of LCS at 12 months after discharge.

RESULTS

Self-reported LCS was detected in 76% of participants at 3 months and in 43% at 12 months after discharge. Patients who reported incomplete recovery at 1 year were characterised by a higher burden of comorbidities (Charlson index of 0.69±0.96 vs 0.31±0.51, p=0.001) and residual pulmonary consolidations (1.56±1.78 vs 0.98±1.56, p=0.034), worse blood pressure (BP) control (systolic BP of 138.1±16.2 vs 132.2±15.8 mm Hg, p=0.041), renal (estimated glomerular filtration rate of 59.5±14.7 vs 69.8±20.7 mL/min/1.73 m2, p=0.007) and endothelial function (flow-mediated dilation of the brachial artery of 10.4±5.4 vs 12.4±5.6%, p=0.048), higher in-hospital levels of liver enzymes (alanine aminotransferase (ALT) of 76.3±60.8 vs 46.3±25.3 IU/L, p=0.002) and erythrocyte sedimentation rate (ESR) (34.3±12.1 vs 28.3±12.6 mm/h, p=0.008), slightly higher indices of ventricular longitudinal function (left ventricular (LV) global longitudinal strain (GLS) of 18.0±2.4 vs 17.0±2.3%, p=0011) and higher levels of Hospital Anxiety and Depression Scale anxiety (7.3±4.2 vs 5.6±3.8, p=0.011) and depression scores (6.4±3.9 vs 4.9±4.3, p=0.022) and EFTER-COVID study physical symptoms score (12.3±3.8 vs 9.2±4.2, p<0.001). At 1 month postdischarge, the persisting differences included marginally higher LV GLS, mitral E/e' ratio and significantly higher levels of both resting and exertional physical symptoms versus patients who reported complete recovery. Logistic regression and machine learning-based binary classification models have been developed to predict the persistence of LCS symptoms at 12 months after discharge.

CONCLUSIONS

Compared with post-COVID-19 patients who have completely recovered by 12 months after hospital discharge, those who have subsequently developed 'very long' COVID were characterised by a variety of more pronounced residual predischarge abnormalities that had mostly subsided by 1 month, except for steady differences in the physical symptoms levels. A simple artificial neural networks-based binary classification model using peak ESR, creatinine, ALT and weight loss during the acute phase, predischarge 6-minute walk distance and complex survey-based symptoms assessment as inputs has shown a 92% accuracy with an area under receiver-operator characteristic curve 0.931 in prediction of LCS symptoms persistence at 12 months.

Diagnostic Power of MicroRNAs in Melanoma: Integrating Machine Learning for Enhanced Accuracy and Pathway Analysis

This study identifies microRNAs (miRNAs) with significant discriminatory power in distinguishing melanoma from nevus, notably hsa-miR-26a and hsa-miR-211, which have exhibited diagnostic potential with accuracy of 81% and 78% respectively. To enhance diagnostic accuracy, we integrated miRNAs into various machine-learning (ML) models. Incorporating miRNAs with AUC scores above 0.70 significantly improved diagnostic accuracy to 94%, with a sensitivity of 91%. These findings underscore the potential of ML models to leverage miRNA data for enhanced melanoma diagnosis. Additionally, using the miRNet tool, we constructed a network of miRNA-miRNA interactions, revealing 170 key genes in melanoma pathophysiology. Protein-protein interaction network analysis via Cytoscape identified hub genes including MYC, BRCA1, JUN, AURKB, CDKN2A, DDX5, MAPK14, DDX3X, DDX6, FOXM1 and GSK3B. The identification of hub genes and their interactions with miRNAs enhances our understanding of the molecular mechanisms driving melanoma. Pathway enrichment analyses highlighted key pathways associated with differentially expressed miRNAs, including the PI3K/AKT, TGF-beta signalling pathway and cell cycle regulation. These pathways are implicated in melanoma development and progression, reinforcing the significance of our findings. The functional enrichment of miRNAs suggests their critical role in modulating essential pathways in melanoma, suggesting their potential as therapeutic targets.

Using GWAS and Machine Learning to Identify and Predict Genetic Variants Associated with Foodborne Bacteria Phenotypic Traits

One of the main challenges in food microbiology is to prevent the risk of outbreaks by avoiding the distribution of food contaminated by bacteria. This requires constant monitoring of the circulating strains throughout the food production chain. Bacterial genomes contain signatures of natural evolution and adaptive markers that can be exploited to better understand the behavior of pathogen in the food industry. The monitoring of foodborne strains can therefore be facilitated by the use of these genomic markers capable of rapidly providing essential information on isolated strains, such as the source of contamination, risk of illness, potential for biofilm formation, and tolerance or resistance to biocides. The increasing availability of large genome datasets is enhancing the understanding of the genetic basis of complex traits such as host adaptation, virulence, and persistence. Genome-wide association studies have shown very promising results in the discovery of genomic markers that can be integrated into rapid detection tools. In addition, machine learning has successfully predicted phenotypes and classified important traits. Genome-wide association and machine learning tools have therefore the potential to support decision-making circuits intending at reducing the burden of foodborne diseases. The aim of this chapter review is to provide knowledge on the use of these two methods in food microbiology and to recommend their use in the field.

Utilising AI technique to identify depression risk among doctoral students

The phenomenon that the depression risk among doctoral students is higher than that of other groups should not be ignored. Despite this, studies specifically addressing depression risk in doctoral students are relatively scarce, and existing findings are not universally applicable. Using neural network feature extraction technology, this study aims to investigate the factors contributing to the high depression risk of doctoral students and effectively identify doctoral students at depression risk, so as to propose corresponding improvement strategies to prevent and intervene doctoral students with depression risk for universities. Based on the data from the 2019 Nature Global Doctoral Student Survey, we first screened 13 highly relevant features from a total of 37 features potentially related to the risk of depression among doctoral students by Random Forest algorithm. Subsequently, we trained the optimal prediction model to predict the doctoral students with depression risk using a Multilayer Perceptron (MLP), achieving an accuracy of 89.09% on the test set. Additionally, this study constructed a group portrait of doctoral students at risk of depression, and found that overwork, poor work-life balance, and poor supervisor-student relationship, etc., were typical characteristics among these students. Finally, we proposed several improvement strategies for higher education institutions. Our research offers a new perspective on utilising artificial intelligence (AI) methods to tackle educational challenges, particularly in the identification and support of doctoral students at risk of depression, thereby enhancing their mental health.

The Role of Machine Learning in the Detection of Cardiac Fibrosis in Electrocardiograms: Scoping Review

BACKGROUND

Cardiovascular disease remains the leading cause of mortality worldwide. Cardiac fibrosis impacts the underlying pathophysiology of many cardiovascular diseases by altering structural integrity and impairing electrical conduction. Identifying cardiac fibrosis is essential for the prognosis and management of cardiovascular disease; however, current diagnostic methods face challenges due to invasiveness, cost, and inaccessibility. Electrocardiograms (ECGs) are widely available and cost-effective for monitoring cardiac electrical activity. While ECG-based methods for inferring fibrosis exist, they are not commonly used due to accuracy limitations and the need for cardiac expertise. However, the ECG shows promise as a target for machine learning (ML) applications in fibrosis detection.

OBJECTIVE

This study aims to synthesize and critically evaluate the current state of ECG-based ML approaches for cardiac fibrosis detection.

METHODS

We conducted a scoping review of research in ECG-based ML applications to identify cardiac fibrosis. Comprehensive searches were performed in PubMed, IEEE Xplore, Scopus, Web of Science, and DBLP databases, including publications up to October 2024. Studies were included if they applied ML techniques to detect cardiac fibrosis using ECG or vectorcardiogram data and provided sufficient methodological details and outcome metrics. Two reviewers independently assessed eligibility and extracted data on the ML models used, their performance metrics, study designs, and limitations.

RESULTS

We identified 11 studies evaluating ML approaches for detecting cardiac fibrosis using ECG data. These studies used various ML techniques, including classical (8/11, 73%), ensemble (3/11, 27%), and deep learning models (4/11, 36%). Support vector machines were the most used classical model (6/11, 55%), with the best-performing models of each study achieving accuracies of 77% to 93%. Among deep learning approaches, convolutional neural networks showed promising results, with one study reporting an area under the receiver operating characteristic curve (AUC) of 0.89 when combined with clinical features. Notably, a large-scale convolutional neural network study (n=14,052) achieved an AUC of 0.84 for detecting cardiac fibrosis, outperforming cardiologists (AUC 0.63-0.66). However, many studies had limited sample sizes and lacked external validation, potentially impacting the generalizability of the findings. Variability in reporting methods may affect the reproducibility and applicability of these ML-based approaches.

CONCLUSIONS

ML-augmented ECG analysis shows promise for accessible and cost-effective detection of cardiac fibrosis. However, there are common limitations with respect to study design and insufficient external validation, raising concerns about the generalizability and clinical applicability of the findings. Inconsistencies in methodologies and incomplete reporting further impede cross-study comparisons. Future work may benefit from using prospective study designs, larger and more clinically and demographically diverse datasets, advanced ML models, and rigorous external validation. Addressing these challenges could pave the way for the clinical implementation of ML-based ECG detection of cardiac fibrosis to improve patient outcomes and health care resource allocation.

Using Machine Learning to Fight Child Acute Malnutrition and Predict Weight Gain During Outpatient Treatment with a Simplified Combined Protocol

BACKGROUND/OBJECTIVES

Child acute malnutrition is a global public health problem, affecting 45 million children under 5 years of age. The World Health Organization recommends monitoring weight gain weekly as an indicator of the correct treatment. However, simplified protocols that do not record the weight and base diagnosis and follow-up in arm circumference at discharge are being tested in emergency settings. The present study aims to use machine learning techniques to predict weight gain based on the socio-economic characteristics at admission for the children treated under a simplified protocol in the Diffa region of Niger.

METHODS

The sample consists of 535 children aged 6-59 months receiving outpatient treatment for acute malnutrition, for whom information on 51 socio-economic variables was collected. First, the Variable Selection Using Random Forest (VSURF) algorithm was used to select the variables associated with weight gain. Subsequently, the dataset was partitioned into training/testing, and an ensemble model was adjusted using five algorithms for prediction, which were combined using a Random Forest meta-algorithm. Afterward, Receiver Operating Characteristic (ROC) curves were used to identify the optimal cut-off point for predicting the group of individuals most vulnerable to developing low weight gain.

RESULTS

The critical variables that influence weight gain are water, hygiene and sanitation, the caregiver's employment-socio-economic level and access to treatment. The final ensemble prediction model achieved a better fit (R2 = 0.55) with respect to the individual algorithms (R2 = 0.14-0.27). An optimal cut-off point was identified to establish low weight gain, with an Area Under the Curve (AUC) of 0.777 at a value of <6.5 g/kg/day. The ensemble model achieved a success rate of 84% (78/93) at the identification of individuals below <6.5 g/kg/day in the test set.

CONCLUSIONS

The results highlight the importance of adapting the cut-off points for weight gain to each context, as well as the practical usefulness that these techniques can have in optimizing and adapting to the treatment in humanitarian settings.

Machine Learning and Metabolomics Predict Mesenchymal Stem Cell Osteogenic Differentiation in 2D and 3D Cultures

Stem cells have been widely used to produce artificial bone grafts. Nonetheless, the variability in the degree of stem cell differentiation is an inherent drawback of artificial graft development and requires robust evaluation tools that can certify the quality of stem cell-based products and avoid source-tissue-related and patient-specific variability in outcomes. Omics analyses have been utilised for the evaluation of stem cell attributes in all stages of stem cell biomanufacturing. Herein, metabolomics in combination with machine learning was utilised for the benchmarking of osteogenic differentiation quality in 2D and 3D cultures. Metabolomics analysis was performed with the use of gas chromatography-mass spectrometry (GC-MS). A set of 11 metabolites was used to train an XGboost model which achieved excellent performance in distinguishing between differentiated and undifferentiated umbilical cord blood mesenchymal stem cells (UCB MSCs). The model was benchmarked against samples not present in the training set, being able to efficiently capture osteogenesis in 3D UCB MSC cultures with an area under the curve (AUC) of 82.6%. On the contrary, the model did not capture any differentiation in Wharton's Jelly MSC samples, which are well-known underperformers in osteogenic differentiation (AUC of 56.2%). Mineralisation was significantly correlated with the levels of fumarate, glycerol, and myo-inositol, the four metabolites found most important for model performance (R2 = 0.89, R2 = 0.94, and R2 = 0.96, and p = 0.016, p = 0.0059, and p = 0.0022, respectively). In conclusion, our results indicate that metabolomics in combination with machine learning can be used for the development of reliable potency assays for the evaluation of Advanced Therapy Medicinal Products.

Cytokine profiles as predictors of HIV incidence using machine learning survival models and statistical interpretable techniques

HIV remains a critical global health issue, with an estimated 39.9 million people living with the virus worldwide by the end of 2023 (according to WHO). Although the epidemic's impact varies significantly across regions, Africa remains the most affected. In the past decade, considerable efforts have focused on developing preventive measures, such as vaccines and pre-exposure prophylaxis, to combat sexually transmitted HIV. Recently, cytokine profiles have gained attention as potential predictors of HIV incidence due to their involvement in immune regulation and inflammation, presenting new opportunities to enhance preventative strategies. However, the high-dimensional, time-varying nature of cytokine data collected in clinical research, presents challenges for traditional statistical methods like the Cox proportional hazards (PH) model to effectively analyze survival data related to HIV. Machine learning (ML) survival models offer a robust alternative, especially for addressing the limitations of the PH model's assumptions. In this study, we applied survival support vector machine (SSVM) and random survival forest (RSF) models using changes or means in cytokine levels as predictors to assess their association with HIV incidence, evaluate variable importance, measure predictive accuracy using the concordance index (C-index) and integrated Brier score (IBS) and interpret the model's predictions using Shapley additive explanations (SHAP) values. Our results indicated that RSFs models outperformed SSVMs models, with the difference covariate model performing better than the mean covariate model. The highest C-index for SSVM was 0.7180 under the difference covariate model, while for RSF, it reached 0.8801 under the difference covariate model using the log-rank split rule. Key cytokines identified as positive predictors of HIV incidence included TNF-A, BASIC-FGF, IL-5, MCP-3, and EOTAXIN, while 29 cytokines were negative predictors. Baseline factors such as condom use frequency, treatment status, number of partners, and sexual activity also emerged as significant predictors. This study underscored the potential of cytokine profiles for predicting HIV incidence and highlighted the advantages of RSFs models in analyzing high-dimensional, time-varying data over SSVMs. It further through ablation studies emphasized the importance of selecting key features within mean and difference based covariate models to achieve an optimal balance between model complexity and predictive accuracy.

Development and Validation of a Smartphone Application for Neonatal Jaundice Screening

Importance

This diagnostic study describes the merger of domain knowledge (Kramer principle of dermal advancement of icterus) with current machine learning (ML) techniques to create a novel tool for screening of neonatal jaundice (NNJ), which affects 60% of term and 80% of preterm infants.

Objective

This study aimed to develop and validate a smartphone-based ML app to predict bilirubin (SpB) levels in multiethnic neonates using skin color analysis.

Design, Setting, and Participants

This diagnostic study was conducted between June 2022 and June 2024 at a tertiary hospital and 4 primary-care clinics in Singapore with a consecutive sample of neonates born at 35 or more weeks' gestation and within 21 days of birth.

Exposure

The smartphone-based ML app captured skin images via the central aperture of a standardized color calibration sticker card from multiple regions of interest arranged in a cephalocaudal fashion, following the Kramer principle of dermal advancement of icterus. The ML model underwent iterative development and k-folds cross-validation, with performance assessed based on root mean squared error, Pearson correlation, and agreement with total serum bilirubin (TSB). The final ML model underwent temporal validation.

Main Outcomes and Measures

Linear correlation and statistical agreement between paired SpB and TSB; sensitivity and specificity for detection of TSB equal to or greater than 17mg/dL with SpB equal to or greater than 13 mg/dL were assessed.

Results

The smartphone-based ML app was validated on 546 neonates (median [IQR] gestational age, 38.0 [35.0-41.0] weeks; 286 [52.4%] male; 315 [57.7%] Chinese, 35 [6.4%] Indian, 169 [31.0%] Malay, and 27 [4.9%] other ethnicities). Iterative development and cross-validation was performed on 352 neonates. The final ML model (ensembled gradient boosted trees) incorporated yellowness indicators from the forehead, sternum, and abdomen. Temporal validation on 194 neonates yielded a Pearson r of 0.84 (95% CI, 0.79-0.88; P < .001), 82% of data pairs within clinically acceptable limits of 3 mg/dL, sensitivity of 100%, specificity of 70%, positive predictive value of 10%, negative predictive value of 100%, positive likelihood ratio of 3.3, negative likelihood ratio of 0, and area under the receiver operating characteristic curve of 0.89 (95% CI, 0.82-0.96).

Conclusions and Relevance

In this diagnostic study of a new smartphone-based ML app, there was good correlation and statistical agreement with TSB with sensitivity of 100%. The screening tool has the potential to be an NNJ screening tool, with treatment decisions based on TSB (reference standard). Further prospective studies are needed to establish the generalizability and cost-effectiveness of the screening tool in the clinical setting.

Machine learning-guided differential gene expression analysis identifies a highly-connected seven-gene cluster in triple-negative breast cancer

Background

One of the most challenging cancers is triple-negative breast cancer, which is subdivided into many molecular subtypes. Due to the high degree of heterogeneity, the role of precision medicine remains challenging. With the use of machine learning (ML)-guided gene selection, the differential gene expression analysis can be optimized, and eventually, the process of precision medicine can see great advancement through biomarker discovery.

Purpose

Enhancing precision medicine in the oncology field by identification of the most representative differentially-expressed genes to be used as biomarkers or as novel drug targets.

Methods

By utilizing data from the Gene Expression Omnibus (GEO) repository and The Cancer Genome Atlas (TCGA), we identified the differentially expressed genes using the linear model for microarray analysis (LIMMA) and edgeR algorithms, and applied ML-based feature selection using several algorithms.

Results

A total of 27 genes were selected by merging features identified with both LIMMA and ML-based feature selection methods. The models with the highest area under the curve (AUC) are CatBoost, Extreme Gradient Boosting (XGBoost), Random Forest, and Multi-Layer Perceptron classifiers. ESR1, FOXA1, GATA3, XBP1, GREB1, AR, and AGR2 were identified as hub genes in a highly interconnected cluster.

Conclusion

ML-based gene selection shows a great impact on the identification of hub genes. The ML models built can improve precision oncology in diagnosis and prognosis. The identified hub genes can serve as biomarkers and warrant further research for potential drug target development.

Do machine learning methods improve prediction of ambient air pollutants with high spatial contrast? A systematic review

BACKGROUND & OBJECTIVE

The use of machine learning for air pollution modelling is rapidly increasing. We conducted a systematic review of studies comparing statistical and machine learning models predicting the spatiotemporal variation of ambient nitrogen dioxide (NO2), ultrafine particles (UFPs) and black carbon (BC) to determine whether and in which scenarios machine learning generates more accurate predictions.

METHODS

Web of Science and Scopus were searched up to June 13, 2024. All records were screened by two independent reviewers. Differences in the coefficient of determination (R2) and Root Mean Square Error (RMSE) between best statistical and machine learning methods were compared across categories of methodological elements.

RESULTS

A total of 38 studies with 46 model comparisons (30 for NO2, 8 for UFPs and 8 for BC) were included. Linear non-regularized methods and Random Forest were most frequently used. Machine learning outperformed statistical models in 34 comparisons. Mean differences (95% confidence intervals) in R2 and RMSE between best machine learning and statistical models were 0.12 (0.08, 0.17) and 20% (11%, 29%) respectively. Tree-based methods performed best in 12 of 17 multi-model comparisons. Nonlinear or regularization regression methods were used in only 12 comparisons and provided similar performance to machine learning methods.

CONCLUSION

This systematic review suggests that machine learning methods, especially tree-based methods, may be superior to linear non-regularized methods for predicting ambient concentrations of NO2, UFPs and BC. Additional comparison studies using nonlinear, regularized and a wider array of machine learning methods are needed to confirm their relative performance. Future air pollution studies would also benefit from more explicit and standardized reporting of methodologies and results.

A beginner's guide to supervised analysis for mass cytometry data in cancer biology

Mass cytometry enables deep profiling of biological samples at single-cell resolution. This technology is more than relevant in cancer research due to high cellular heterogeneity and complexity. Downstream analysis of high-dimensional datasets increasingly relies on machine learning (ML) to extract clinically relevant information, including supervised algorithms for classification and regression purposes. In cancer research, they are used to develop predictive models that will guide clinical decision making. However, the development of supervised algorithms faces major challenges, such as sufficient validation, before being translated into the clinics. In this work, we provide a framework for the analysis of mass cytometry data with a specific focus on supervised algorithms and practical examples of their applications. We also raise awareness on key issues regarding good practices for researchers curious to implement supervised ML on their mass cytometry data. Finally, we discuss the challenges of supervised ML application to cancer research.

Enhancing mortality prediction in patients with spontaneous intracerebral hemorrhage: Radiomics and supervised machine learning on non-contrast computed tomography

Purpose

This study aims to develop a Radiomics-based Supervised Machine-Learning model to predict mortality in patients with spontaneous intracerebral hemorrhage (sICH).

Methods

Retrospective analysis of a prospectively collected clinical registry of patients with sICH consecutively admitted at a single academic comprehensive stroke center between January-2016 and April-2018. We conducted an in-depth analysis of 105 radiomic features extracted from 105 patients. Following the identification and handling of missing values, radiomics values were scaled to 0-1 to train different classifiers. The sample was split into 80-20 % training-test and validation cohort in a stratified fashion. Random Forest(RF), K-Nearest Neighbor(KNN), and Support Vector Machine(SVM) classifiers were evaluated, along with several feature selection methods and hyperparameter optimization strategies, to classify the binary outcome of mortality or survival during hospital admission. A tenfold stratified cross-validation method was used to train the models, and average metrics were calculated.

Results

RF, KNN, and SVM, with the "DropOut+SelectKBest" feature selection strategy and no hyperparameter optimization, demonstrated the best performances with the least number of radiomic features and the most simplified models, achieving a sensitivity range between 0.90 and 0.95 and AUC range from 0.97 to 1 on the validation dataset. Regarding the confusion matrix, the SVM model did not predict any false negative test (negative predicted value 1).

Conclusion

Radiomics-based Supervised Machine Learning models can predict mortality during admission in patients with sICH. SVM with the "DropOut+SelectKBest" feature selection strategy and no hyperparameter optimization was the best simplified model to detect mortality during admission in patients with sICH.

FRELSA: A dataset for frailty in elderly people originated from ELSA and evaluated through machine learning models

BACKGROUND

Frailty is an age-related syndrome characterized by loss of strength and exhaustion and associated with multi-morbidity. Early detection and prediction of the appearance of frailty could help older people age better and prevent them from needing invasive and expensive treatments. Machine learning techniques show promising results in creating a medical support tool for such a task.

METHODS

This study aims to create a dataset for machine learning-based frailty studies, using Fried's Frailty Phenotype definition. Starting from a longitudinal study on aging in the UK population, we defined a frailty label for each subject. We evaluated the definition by training seven different models for detecting frailty with data that were contemporary to the ones used for the definition. We then integrated more data from two years before to obtain prediction models with a 24-month horizon. Features selection was performed using the MultiSURF algorithm, which ranks all features in order of relevance to the detection or prediction task.

RESULTS

We present a new frailty dataset of 5303 subjects and more than 6500 available features. It is publicly available, provided one has access to the original English Longitudinal Study of Ageing dataset. The dataset is balanced after grouping frailty with pre-frailty, and it is suitable for multiclass or binary classification and prediction problems. The seven tested architectures performed similarly, forming a solid baseline that can be improved with future work. Linear regression achieved the best F-score and AUROC in detection and prediction tasks.

CONCLUSIONS

Creating new frailty-annotated datasets of this size is necessary to develop and improve the frailty prediction techniques. We have shown that our dataset can be used to study and test machine learning models to detect and predict frailty. Future work should improve models' architecture and performance, consider explainability, and possibly enrich the dataset with older waves.

The development of a C5.0 machine learning model in a limited data set to predict early mortality in patients with ARDS undergoing an initial session of prone positioning

BACKGROUND

Acute Respiratory Distress Syndrome (ARDS) has a high morbidity and mortality. One therapy that can decrease mortality is ventilation in the prone position (PP). Patients undergoing PP are amongst the sickest, and there is a need for early identification of patients at particularly high risk of death. These patients may benefit from an in-depth review of treatment or consideration of rescue therapies. We report the development of a machine learning model trained to predict early mortality in patients undergoing prone positioning as part of the management of their ARDS.

METHODS

Prospectively collected clinical data were analysed retrospectively from a single tertiary ICU. The records of patients who underwent an initial session of prone positioning whilst receiving invasive mechanical ventilation were identified (n = 131). The decision to perform prone positioning was based on the criteria in the PROSEVA study. A C5.0 classifier algorithm with adaptive boosting was trained on data gathered before, during, and after initial proning. Data was split between training (85% of data) and testing (15% of data). Hyperparameter tuning was achieved through a grid-search using a maximal entropy configuration. Predictions for 7-day mortality after initial proning session were made on the training and testing data.

RESULTS

The model demonstrated good performance in predicting 7-day mortality (AUROC: 0.89 training, 0.78 testing). Seven variables were used for prediction. Sensitivity was 0.80 and specificity was 0.67 on the testing data set. Patients predicted to survive had 13.3% mortality, while those predicted to die had 66.67% mortality. Among patients in whom the model predicted patient would survive to day 7 based on their response, mortality at day 7 was 13.3%. Conversely, if the model predicted the patient would not survive to day 7, mortality was 66.67%.

CONCLUSIONS

This proof-of-concept study shows that with a limited data set, a C5.0 classifier can predict 7-day mortality from a number of variables, including the response to initial proning, and identify a cohort at significantly higher risk of death. This can help identify patients failing conventional therapies who may benefit from a thorough review of their management, including consideration of rescue treatments, such as extracorporeal membrane oxygenation. This study shows the potential of a machine learning model to identify ARDS patients at high risk of early mortality following PP. This information can guide clinicians in tailoring treatment strategies and considering rescue therapies. Further validation in larger cohorts is needed.

Machine Learning Model for Predicting Walking Ability in Lower Limb Amputees

Background/Objectives: The number of individuals with lower limb loss (LLL) is rising. Therefore, identifying the walking potential in individuals with LLL and prescribing adequate prosthetic systems are crucial. Various factors can influence participants' walking ability, to different extents. The aim of the present study was to apply machine learning methods to develop a predictive mode. This model can assist rehabilitation and limb loss care teams in making informed decisions regarding prosthesis prescription and predicting walking ability in individuals with LLL. Methods: The present study was designed as a prospective cross-sectional study encompassing 104 consecutively recruited participants with LLL (average age 62.1 ± 10.9 years, 80 (76.9%) men) at the Medical Rehabilitation Clinic. Demographic, physical, psychological, and social status data of patients were collected at the beginning of the rehabilitation program. At the end of the treatment, K-level estimation of functional ability, a Timed Up and Go Test (TUG), and a Two-Minute Walking Test (TMWT) were performed. Support vector machines (SVM) were used to develop the prediction model. Results: Three decision trees were created, one for each output, as follows: K-level, TUG, and TMWT. For all three outputs, there were eight significant predictors (balance, body mass index, age, Beck depression inventory, amputation level, muscle strength of the residual extremity hip extensors, intact extremity (IE) plantar flexors, and IE hip extensors). For the K-level, the ninth predictor was The Multidimensional Scale of Perceived Social Support (MSPSS). Conclusions: Using the SVM model, we can predict the K-level, TUG, and TMWT with high accuracy. These clinical assessments could be incorporated into routine clinical practice to guide clinicians and inform patients of their potential level of ambulation.

Sex estimation from patellar measurements in a contemporary Italian population: a machine learning approach

Biological sex estimation in forensic anthropology is a crucial topic, and the patella has shown promise in this regard due to its sexual dimorphism. This study uses 12 machine learning models for sex estimation based on three patellar measurements (maximum height, breadth, and thickness). Data was collected from 180 skeletons of a contemporary Italian population (83 males and 97 females) as well as from an independent sample of 21 forensic cases (13 males and 8 females). Statistical analyses indicated that each of the variables exhibited significant sexual dimorphism. To predict biological sex, the classifiers were built using 70% of a reference sample, then tested on the remaining 30% of the original sample and then tested again on the independent sample. The different classifiers generated accuracies varied between 0.85 and 0.91 on the reference sample and between 0.71 and 0.95 for the validation sample. SVM classifier stood out with the highest accuracy and seemed the best model for our study.This study contributes to the growing application of machine learning in forensic anthropology by being the first to apply such techniques to patellar measurements in an Italian population. It aims to enhance the accuracy and efficiency of biological sex estimation from the patella, building on promising results observed with other skeletal elements.

Artificial intelligence-based motion tracking in cancer radiotherapy: A review

Radiotherapy aims to deliver a prescribed dose to the tumor while sparing neighboring organs at risk (OARs). Increasingly complex treatment techniques such as volumetric modulated arc therapy (VMAT), stereotactic radiosurgery (SRS), stereotactic body radiotherapy (SBRT), and proton therapy have been developed to deliver doses more precisely to the target. While such technologies have improved dose delivery, the implementation of intra-fraction motion management to verify tumor position at the time of treatment has become increasingly relevant. Artificial intelligence (AI) has recently demonstrated great potential for real-time tracking of tumors during treatment. However, AI-based motion management faces several challenges, including bias in training data, poor transparency, difficult data collection, complex workflows and quality assurance, and limited sample sizes. This review presents the AI algorithms used for chest, abdomen, and pelvic tumor motion management/tracking for radiotherapy and provides a literature summary on the topic. We will also discuss the limitations of these AI-based studies and propose potential improvements.

A novel open-access artificial-intelligence-driven platform for CNS drug discovery utilizing adult zebrafish

BACKGROUND

Although zebrafish are increasingly utilized in biomedicine for CNS disease modelling and drug discovery, this generates big data necessitating objective, precise and reproducible analyses. The artificial intelligence (AI) applications have empowered automated image recognition and video-tracking to ensure more efficient behavioral testing.

NEW METHOD

Capitalizing on several AI tools that most recently became available, here we present a novel open-access AI-driven platform to analyze tracks of adult zebrafish collected from in vivo neuropharmacological experiments. For this, we trained the AI system to distinguish zebrafish behavioral patterns following systemic treatment with several well-studied psychoactive drugs - nicotine, caffeine and ethanol.

RESULTS

Experiment 1 showed the ability of the AI system to distinguish nicotine and caffeine with 75 % and ethanol with 88 % probability and high (81 %) accuracy following a post-training exposure to these drugs. Experiment 2 further validated our system with additional, previously unexposed compounds (cholinergic arecoline and varenicline, and serotonergic fluoxetine), used as positive and negative controls, respectively.

COMPARISON WITH EXISTING METHODS

The present study introduces a novel open-access AI-driven approach to analyze locomotor activity of adult zebrafish.

CONCLUSIONS

Taken together, these findings support the value of custom-made AI tools for unlocking full potential of zebrafish CNS drug research by monitoring, processing and interpreting the results of in vivo experiments.

European cross-cultural neuropsychological test battery (CNTB) for the assessment of cognitive impairment in multiple sclerosis: Cognitive phenotyping and classification supported by machine learning techniques

BACKGROUND

The European Cross-Cultural Neuropsychological Test Battery (CNTB) has been proposed as a comprehensive battery for cognitive assessment, reducing the potential impact of cultural variables. In this validation study, we aimed to evaluate the diagnostic capacity of CNTB for the assessment of participants with multiple sclerosis (pwMS) compared to the Neuronorma battery (NN) according to the International Classification of Cognitive Disorders in MS criteria, and to develop machine learning (ML) algorithms to improve the diagnostic capacity of CNTB and to select the most relevant tests.

METHODS

Sixty pwMS and 60 healthy controls (HC) with no differences in sex, age, or years of education were enrolled. All participants completed the CNTB and pwMS were also examined with NN, depression, and fatigue scales. Impaired domains and cognitive phenotypes were defined following ICCoDiMS based on CNTB scores and compared to NN, according to -1SD and -1.5SD cutoff scores. To select the most relevant tests, random forest (RF) was performed for different binary classifications.

RESULTS

PwMS showed a lower performance compared to HC with medium-large effect sizes, in episodic memory, executive function, attention, and processing speed, in accordance with their characteristic cognitive profile. There were no differences in impaired domains or cognitive phenotypes between CNTB and NN, highlighting the role of episodic memory, executive function, attention, and processing speed tests. The most relevant tests identified by RF were consistent with inter-group comparisons and allowed a better classification than SD cutoff scores.

CONCLUSION

CNTB is a valid test for cognitive diagnosis in pwMS, including key tests for the most frequently impaired cognitive domains in MS. The use of ML techniques may also be useful to improve diagnosis, especially in some tests with lower sensitivity.

Non-invasive prediction of human embryonic ploidy using artificial intelligence: a systematic review and meta-analysis

Background

Embryonic ploidy is critical for the success of embryo transfer. Currently, preimplantation genetic testing for aneuploidy (PGT-A) is the gold standard for detecting ploidy abnormalities. However, PGT-A has several inherent limitations, including invasive biopsy, high economic burden, and ethical constraints. This paper provides the first comprehensive systematic review and meta-analysis of the performance of artificial intelligence (AI) algorithms using embryonic images for non-invasive prediction of embryonic ploidy.

Methods

Comprehensive searches of studies that developed or utilized AI algorithms to predict embryonic ploidy from embryonic imaging, published up until August 10, 2024, across PubMed, MEDLINE, Embase, IEEE, SCOPUS, Web of Science, and the Cochrane Central Register of Controlled Trials were performed. Studies with prospective or retrospective designs were included without language restrictions. The summary receiver operating characteristic curve, along with pooled sensitivity and specificity, was estimated using a bivariate random-effects model. The risk of bias and study quality were evaluated using the QUADAS-AI tool. Heterogeneity was quantified using the inconsistency index (I 2 ), derived from Cochran's Q test. Predefined subgroup analyses and bivariate meta-regression were conducted to explore potential sources of heterogeneity. This study was registered with PROSPERO (CRD42024500409).

Findings

Twenty eligible studies were identified, with twelve studies included in the meta-analysis. The pooled sensitivity, specificity, and area under the curve of AI for predicting embryonic euploidy were 0.71 (95% CI: 0.59-0.81), 0.75 (95% CI: 0.69-0.80), and 0.80 (95% CI: 0.76-0.83), respectively, based on a total of 6879 embryos (3110 euploid and 3769 aneuploid). Meta-regression and subgroup analyses identified the type of AI-driven decision support system, external validation, risk of bias, and year of publication as the primary contributors to the observed heterogeneity. There was no evidence of publication bias.

Interpretation

Our findings indicate that AI algorithms exhibit promising performance in predicting embryonic euploidy based on embryonic imaging. Although the current AI models developed cannot entirely replace invasive methods for determining embryo ploidy, AI demonstrates promise as an auxiliary decision-making tool for embryo selection, particularly for individuals who are unable to undergo PGT-A. To enhance the quality of future research, it is essential to overcome the specific challenges and limitations associated with AI studies in reproductive medicine.

Funding

This work was supported by the National Key R&D Program of China (2022YFC2702905), the Shengjing Freelance Researcher Plan of Shengjing Hospital and the 345 talent project of Shengjing Hospital.

Recommendations for the creation of benchmark datasets for reproducible artificial intelligence in radiology

Various healthcare domains have witnessed successful preliminary implementation of artificial intelligence (AI) solutions, including radiology, though limited generalizability hinders their widespread adoption. Currently, most research groups and industry have limited access to the data needed for external validation studies. The creation and accessibility of benchmark datasets to validate such solutions represents a critical step towards generalizability, for which an array of aspects ranging from preprocessing to regulatory issues and biostatistical principles come into play. In this article, the authors provide recommendations for the creation of benchmark datasets in radiology, explain current limitations in this realm, and explore potential new approaches. CLINICAL RELEVANCE STATEMENT: Benchmark datasets, facilitating validation of AI software performance can contribute to the adoption of AI in clinical practice. KEY POINTS: Benchmark datasets are essential for the validation of AI software performance. Factors like image quality and representativeness of cases should be considered. Benchmark datasets can help adoption by increasing the trustworthiness and robustness of AI.

Dental Composite Performance Prediction Using Artificial Intelligence

Objective

There is a need to increase the performance and longevity of dental composites and accelerate the translation of novel composites to the market. This study explores artificial intelligence (AI), specifically machine learning (ML), to predict the performance outcomes (POs) of dental composites from their composite attributes (CAs).

Methods

An extensive dataset from over 200 publications was built and refined to 233 samples with 17 CAs and 7 POs. Nine ML models were evaluated for PO prediction performance using classified data, and Five ML models were evaluated for PO regression analysis.

Results

The KNN model excelled in predicting flexural modulus (FlexMod), Decision Tree model in flexural strength (FlexStr) and volumetric shrinkage (ShrinkV), and Logistic Regression and SVM models in shrinkage stress (ShrinkStr). Receiver operating characteristic area under the curve (ROC AUC) analysis confirmed these results but found that Random Forest was more effective for FlexStr and ShrinkV, suggesting the possibility of Decision Tree overfitting the data. Regression analysis revealed that the Voting Regressor was superior for FlexMod and ShrinkV predictions, while Decision Tree Regression was optimal for FlexStr and ShrinkStr. Feature importance analysis indicated TEGDMA is a key contributor to FlexMod and ShrinkV, BisGMA and UDMA to FlexStr, and depth of cure, degree of monomer-to-polymer conversion, and filler loading to ShrinkStr.

Significance

There is a need to conduct a full analysis using multiple ML models because different models predict different POs better, and for a large, comprehensive dataset to train robust AI models to facilitate the prediction and optimization of composite properties and support the development of new dental materials.

Early prediction of mortality and morbidities in VLBW preterm neonates using machine learning

BACKGROUND

Predicting mortality and specific morbidities before they occur may allow for interventions that may improve health trajectories.

HYPOTHESIS

Integrating key maternal and postnatal infant variables in the first 2 weeks of age into machine learning (ML) algorithms will reliably predict survival and specific morbidities in VLBW preterm infants.

METHODS

ML algorithms were developed to integrate 47 features for predicting mortality, bronchopulmonary dysplasia (BPD), neonatal sepsis, necrotizing enterocolitis (NEC), intraventricular hemorrhage (IVH), cystic periventricular leukomalacia (PVL), and retinopathy of prematurity (ROP). A retrospective cohort (n = 3341) was used to train and validate the models with a repeated 10-fold cross-validation strategy. These models were then tested on a separate cohort (n = 447) to evaluate the final model performance.

RESULTS

Among the seven ML algorithms employed, tree-based ensemble models, specifically Random Forest (RF) and XGBoost, had the best performance metrics. The area under the receiver operating characteristic curve (AUROC) of sepsis with or without meningitis (0.73), NEC (0.73), BPD (0.71), and mortality (0.74) exceeded 0.7, while the area under Precision-Recall curve (AUPRC) for all outcomes was greater than the prevalence, demonstrating effective risk stratification in VLBW preterm infants.

CONCLUSIONS

Our study demonstrates the potential of predictive analytics leveraging ML techniques in advancing precision medicine.

IMPACT

Reliable prediction of adverse outcomes before they occur has the potential to institute interventions and possibly improve health trajectories in VLBW preterm infants. We used machine learning to develop and test predictive models for mortality and five major morbidities in VLBW preterm infants. Individualized prediction of outcomes and individualized interventions will advance Precision Medicine in Neonatology.

Machine learning applications in preventive healthcare: A systematic literature review on predictive analytics of disease comorbidity from multiple perspectives

Artificial intelligence is constantly revolutionizing biomedical research and healthcare management. Disease comorbidity is a major threat to the quality of life for susceptible groups, especially middle-aged and elderly patients. The presence of multiple chronic diseases makes precision diagnosis challenging to realize and imposes a heavy burden on the healthcare system and economy. Given an enormous amount of accumulated health data, machine learning techniques show their capability in handling this puzzle. The present study conducts a review to uncover current research efforts in applying these methods to understanding comorbidity mechanisms and making clinical predictions considering these complex patterns. A descriptive metadata analysis of 791 unique publications aims to capture the overall research progression between January 2012 and June 2023. To delve into comorbidity-focused research, 61 of these scientific papers are systematically assessed. Four predictive analytics of tasks are detected: disease comorbidity data extraction, clustering, network, and risk prediction. It is observed that some machine learning-driven applications address inherent data deficiencies in healthcare datasets and provide a model interpretation that identifies significant risk factors of comorbidity development. Based on insights, both technical and practical, gained from relevant literature, this study intends to guide future interests in comorbidity research and draw conclusions about chronic disease prevention and diagnosis with managerial implications.

Machine learning classification of new firearm injury encounters in the St Louis region: 2010-2020

OBJECTIVES

To improve firearm injury encounter classification (new vs follow-up) using machine learning (ML) and compare our ML model to other common approaches.

MATERIALS AND METHODS

This retrospective study used data from the St Louis region-wide hospital-based violence intervention program data repository (2010-2020). We randomly selected 500 patients with a firearm injury diagnosis for inclusion, with 808 total firearm injury encounters split (70/30) for training and testing. We trained a least absolute shrinkage and selection operator (LASSO) regression model with the following predictors: admission type, time between firearm injury visits, number of prior firearm injury emergency department (ED) visits, encounter type (ED or other), and diagnostic codes. Our gold standard for new firearm injury encounter classification was manual chart review. We then used our test data to compare the performance of our ML model to other commonly used approaches (proxy measures of ED visits and time between firearm injury encounters, and diagnostic code encounter type designation [initial vs subsequent or sequela]). Performance metrics included area under the curve (AUC), sensitivity, and specificity with 95% confidence intervals (CIs).

RESULTS

The ML model had excellent discrimination (0.92, 0.88-0.96) with high sensitivity (0.95, 0.90-0.98) and specificity (0.89, 0.81-0.95). AUC was significantly higher than time-based outcomes, sensitivity was slightly (but not significantly) lower than other approaches, and specificity was higher than all other methods.

DISCUSSION

ML successfully delineated new firearm injury encounters, outperforming other approaches in ruling out encounters for follow-up.

CONCLUSION

ML can be used to identify new firearm injury encounters and may be particularly useful in studies assessing re-injuries.

Identification of diabetic retinopathy classification using machine learning algorithms on clinical data and optical coherence tomography angiography

BACKGROUND

To apply machine learning (ML) algorithms to perform multiclass diabetic retinopathy (DR) classification using both clinical data and optical coherence tomography angiography (OCTA).

METHODS

In this cross-sectional observational study, clinical data and OCTA parameters from 203 diabetic patients (203 eye) were used to establish the ML models, and those from 169 diabetic patients (169 eye) were used for independent external validation. The random forest, gradient boosting machine (GBM), deep learning and logistic regression algorithms were used to identify the presence of DR, referable DR (RDR) and vision-threatening DR (VTDR). Four different variable patterns based on clinical data and OCTA variables were examined. The algorithms' performance were evaluated using receiver operating characteristic curves and the area under the curve (AUC) was used to assess predictive accuracy.

RESULTS

The random forest algorithm on OCTA+clinical data-based variables and OCTA+non-laboratory factor-based variables provided the higher AUC values for DR, RDR and VTDR. The GBM algorithm produced similar results, albeit with slightly lower AUC values. Leading predictors of DR status included vessel density, retinal thickness and GCC thickness, as well as the body mass index, waist-to-hip ratio and glucose-lowering treatment.

CONCLUSIONS

ML-based multiclass DR classification using OCTA and clinical data can provide reliable assistance for screening, referral, and management DR populations.

Trial Factors Associated With Completion of Clinical Trials Evaluating AI: Retrospective Case-Control Study

BACKGROUND

Evaluation of artificial intelligence (AI) tools in clinical trials remains the gold standard for translation into clinical settings. However, design factors associated with successful trial completion and the common reasons for trial failure are unknown.

OBJECTIVE

This study aims to compare trial design factors of complete and incomplete clinical trials testing AI tools. We conducted a case-control study of complete (n=485) and incomplete (n=51) clinical trials that evaluated AI as an intervention of ClinicalTrials.gov.

METHODS

Trial design factors, including area of clinical application, intended use population, and intended role of AI, were extracted. Trials that did not evaluate AI as an intervention and active trials were excluded. The assessed trial design factors related to AI interventions included the domain of clinical application related to organ systems; intended use population for patients or health care providers; and the role of AI for different applications in patient-facing clinical workflows, such as diagnosis, screening, and treatment. In addition, we also assessed general trial design factors including study type, allocation, intervention model, masking, age, sex, funder, continent, length of time, sample size, number of enrollment sites, and study start year. The main outcome was the completion of the clinical trial. Odds ratio (OR) and 95% CI values were calculated for all trial design factors using propensity-matched, multivariable logistic regression.

RESULTS

We queried ClinicalTrials.gov on December 23, 2023, using AI keywords to identify complete and incomplete trials testing AI technologies as a primary intervention, yielding 485 complete and 51 incomplete trials for inclusion in this study. Our nested propensity-matched, case-control results suggest that trials conducted in Europe were significantly associated with trial completion when compared with North American trials (OR 2.85, 95% CI 1.14-7.10; P=.03), and the trial sample size was positively associated with trial completion (OR 1.00, 95% CI 1.00-1.00; P=.02).

CONCLUSIONS

Our case-control study is one of the first to identify trial design factors associated with completion of AI trials and catalog study-reported reasons for AI trial failure. We observed that trial design factors positively associated with trial completion include trials conducted in Europe and sample size. Given the promising clinical use of AI tools in health care, our results suggest that future translational research should prioritize addressing the design factors of AI clinical trials associated with trial incompletion and common reasons for study failure.

Machine learning-based evaluation of prognostic factors for mortality and relapse in patients with acute lymphoblastic leukemia: a comparative simulation study

BACKGROUND

Predicting mortality and relapse in children with acute lymphoblastic leukemia (ALL) is crucial for effective treatment and follow-up management. ALL is a common and deadly childhood cancer that often relapses after remission. In this study, we aimed to apply and evaluate machine learning-based models for predicting mortality and relapse in pediatric ALL patients.

METHODS

This retrospective cohort study was conducted on 161 children aged less than 16 years with ALL. Survival status (dead/alive) and patient experience of relapse (yes/no) were considered as the outcome variables. Ten machine learning (ML) algorithms were used to predict mortality and relapse. The performance of the algorithms was evaluated by cross-validation and reported as mean sensitivity, specificity, accuracy and area under the curve (AUC). Finally, prognostic factors were identified based on the best algorithms.

RESULTS

The mean accuracy of the ML algorithms for prediction of patient mortality ranged from 64 to 74% and for prediction of relapse, it varied from 64 to 84% on test data sets. The mean AUC of the ML algorithms for mortality and relapse was above 64%. The most important prognostic factors for predicting both mortality and relapse were identified as age at diagnosis, hemoglobin and platelets. In addition, significant prognostic factors for predicting mortality included clinical side effects such as splenomegaly, hepatomegaly and lymphadenopathy.

CONCLUSIONS

Our results showed that artificial neural networks and bagging algorithms outperformed other algorithms in predicting mortality, while boosting and random forest algorithms excelled in predicting relapse in ALL patients across all criteria. These results offer significant clinical insights into the prognostic factors for children with ALL, which can inform treatment decisions and improve patient outcomes.

Radiomics of Intrahepatic Cholangiocarcinoma and Peritumoral Tissue Predicts Postoperative Survival: Development of a CT-Based Clinical-Radiomic Model

BACKGROUND

For many tumors, radiomics provided a relevant prognostic contribution. This study tested whether the computed tomography (CT)-based textural features of intrahepatic cholangiocarcinoma (ICC) and peritumoral tissue improve the prediction of survival after resection compared with the standard clinical indices.

METHODS

All consecutive patients affected by ICC who underwent hepatectomy at six high-volume centers (2009-2019) were considered for the study. The arterial and portal phases of CT performed fewer than 60 days before surgery were analyzed. A manual segmentation of the tumor was performed (Tumor-VOI). A 5-mm volume expansion then was applied to identify the peritumoral tissue (Margin-VOI).

RESULTS

The study enrolled 215 patients. After a median follow-up period of 28 months, the overall survival (OS) rate was 57.0%, and the progression-free survival (PFS) rate was 34.9% at 3 years. The clinical predictive model of OS had a C-index of 0.681. The addition of radiomic features led to a progressive improvement of performances (C-index of 0.71, including the portal Tumor-VOI, C-index of 0.752 including the portal Tumor- and Margin-VOI, C-index of 0.764, including all VOIs of the portal and arterial phases). The latter model combined clinical variables (CA19-9 and tumor pattern), tumor indices (density, homogeneity), margin data (kurtosis, compacity, shape), and GLRLM indices. The model had performance equivalent to that of the postoperative clinical model including the pathology data (C-index of 0.765). The same results were observed for PFS.

CONCLUSIONS

The radiomics of ICC and peritumoral tissue extracted from preoperative CT improves the prediction of survival. Both the portal and arterial phases should be considered. Radiomic and clinical data are complementary and achieve a preoperative estimation of prognosis equivalent to that achieved in the postoperative setting.

New horizons in prediction modelling using machine learning in older people's healthcare research

Machine learning (ML) and prediction modelling have become increasingly influential in healthcare, providing critical insights and supporting clinical decisions, particularly in the age of big data. This paper serves as an introductory guide for health researchers and readers interested in prediction modelling and explores how these technologies support clinical decisions, particularly with big data, and covers all aspects of the development, assessment and reporting of a model using ML. The paper starts with the importance of prediction modelling for precision medicine. It outlines different types of prediction and machine learning approaches, including supervised, unsupervised and semi-supervised learning, and provides an overview of popular algorithms for various outcomes and settings. It also introduces key theoretical ML concepts. The importance of data quality, preprocessing and unbiased model performance evaluation is highlighted. Concepts of apparent, internal and external validation will be introduced along with metrics for discrimination and calibration for different types of outcomes. Additionally, the paper addresses model interpretation, fairness and implementation in clinical practice. Finally, the paper provides recommendations for reporting and identifies common pitfalls in prediction modelling and machine learning. The aim of the paper is to help readers understand and critically evaluate research papers that present ML models and to serve as a first guide for developing, assessing and implementing their own.

Machine learning characterization of a rare neurologic disease via electronic health records: a proof-of-principle study on stiff person syndrome

BACKGROUND

Despite the frequent diagnostic delays of rare neurologic diseases (RND), it remains difficult to study RNDs and their comorbidities due to their rarity and hence the statistical underpowering. Affecting one to two in a million annually, stiff person syndrome (SPS) is an RND characterized by painful muscle spasms and rigidity. Leveraging underutilized electronic health records (EHR), this study showcased a machine-learning-based framework to identify clinical features that optimally characterize the diagnosis of SPS.

METHODS

A machine-learning-based feature selection approach was employed on 319 items from the past medical histories of 48 individuals (23 with a diagnosis of SPS and 25 controls) with elevated serum autoantibodies against glutamic-acid-decarboxylase-65 (anti-GAD65) in Dartmouth Health's EHR to determine features with the highest discriminatory power. Each iteration of the algorithm implemented a Support Vector Machine (SVM) model, generating importance scores-SHapley Additive exPlanation (SHAP) values-for each feature and removing one with the least salient. Evaluation metrics were calculated through repeated stratified cross-validation.

RESULTS

Depression, hypothyroidism, GERD, and joint pain were the most characteristic features of SPS. Utilizing these features, the SVM model attained precision of 0.817 (95% CI 0.795-0.840), sensitivity of 0.766 (95% CI 0.743-0.790), F-score of 0.761 (95% CI 0.744-0.778), AUC of 0.808 (95% CI 0.791-0.825), and accuracy of 0.775 (95% CI 0.759-0.790).

CONCLUSIONS

This framework discerned features that, with further research, may help fully characterize the pathologic mechanism of SPS: depression, hypothyroidism, and GERD may respectively represent comorbidities through common inflammatory, genetic, and dysautonomic links. This methodology could address diagnostic challenges in neurology by uncovering latent associations and generating hypotheses for RNDs.