Opening the black box: the promise and limitations of explainable machine learning in cardiology

A latent diffusion approach to visual attribution in medical imaging

Visual attribution in medical imaging seeks to make evident the diagnostically-relevant components of a medical image, in contrast to the more common detection of diseased tissue deployed in standard machine vision pipelines (which are less straightforwardly interpretable/explainable to clinicians). We here present a novel generative visual attribution technique, one that leverages latent diffusion models in combination with domain-specific large language models, in order to generate normal counterparts of abnormal images. The discrepancy between the two hence gives rise to a mapping indicating the diagnostically-relevant image components. To achieve this, we deploy image priors in conjunction with appropriate conditioning mechanisms in order to control the image generative process, including natural language text prompts acquired from medical science and applied radiology. We perform experiments and quantitatively evaluate our results on the COVID-19 Radiography Database containing labelled chest X-rays with differing pathologies via the Frechet Inception Distance (FID), Structural Similarity (SSIM) and Multi Scale Structural Similarity Metric (MS-SSIM) metrics obtained between real and generated images. The resulting system also exhibits a range of latent capabilities including zero-shot localized disease induction, which are evaluated with real examples from the cheXpert dataset.

GDMol: Generative Double-Masking Self-Supervised Learning for Molecular Property Prediction

BACKGROUND

Effective molecular feature representation is crucial for drug property prediction. Recent years have seen increased attention on graph neural networks (GNNs) that are pre-trained using self-supervised learning techniques, aiming to overcome the scarcity of labeled data in molecular property prediction. Traditional GNNs in self-supervised molecular property prediction typically perform a single masking operation on the nodes and edges of the input molecular graph, masking only local information and insufficient for thorough self-supervised training.

METHOD

Hence, we propose a model for molecular property prediction based on generative double-masking self-supervised learning, termed as GDMol. This integrates generative learning into the self-supervised learning framework for latent representation, and applies a second round of masking to these latent representations, enabling the model to better capture global information and semantic knowledge of the molecules for a richer, more informative representation, thereby achieving more accurate and robust molecular property prediction.

RESULTS

Our experiments on 5 datasets demonstrated superior performance of GDMol in predicting molecular properties across different domains. Moreover, we used the masking operation to traverse through the gradient changes of each node, the magnitude and sign of which reflect the positive and negative contribution respectively of the local structure in the molecule to the prediction outcome. This in-depth interpretative analysis not only enhances the model's interpretability, but also provides more targeted insights and direction for optimizing drug molecules.

CONCLUSIONS

In summary, this research offers novel insights on improving molecular property prediction tasks, and paves the way for further research on the application of generative learning and self-supervised learning in the field of chemistry.

Introducing and Validating the Multiphasic Evidential Decision-Making Matrix (MedMax) for Clinical Management in Patients with Intrahepatic Cholangiocarcinoma

Background: Despite the significant advancements of liver surgery in the last few decades, the survival rate of patients with liver and pancreatic cancers has improved by only 10% in 30 years. Precision medicine offers a patient-centered approach, which, when combined with machine learning, could enhance decision making and treatment outcomes in surgical management of ihCC. This study aims to develop a decision support model to optimize treatment strategies for patients with ihCC, a prevalent primary liver cancer. Methods: The decision support model, named MedMax, was developed using three data sources: studies retrieved through a systematic literature review, expert opinions from HPB surgeons, and data from ihCC patients treated at Heidelberg University Hospital. Expert opinions were collected via surveys, with factors rated on a Likert scale, while patient data were used to validate the model's accuracy. Results: The model is structured into four decision-making phases, assessing diagnosis, treatment modality, surgical approach, and prognosis. Prospectively, 44 patients with ihCC were included for internal primary validation of the model. MedMax could predict the appropriate treatment considering the resectability of the lesions in 100% of patients. Also, MedMax could predict a decent surgical approach in 77% of the patients. The model proved effective in making decisions regarding surgery and patient management, demonstrating its potential as a clinical decision support tool. Conclusions: MedMax offers a transparent, personalized approach to decision making in HPB surgery, particularly for ihCC patients. Initial results show high accuracy in treatment selection, and the model's flexibility allows for future expansion to other liver tumors and HPB surgeries. Further validation with larger patient cohorts is required to enhance its clinical utility.

The anesthesiologist's guide to critically assessing machine learning research: a narrative review

Artificial Intelligence (AI), especially Machine Learning (ML), has developed systems capable of performing tasks that require human intelligence. In anesthesiology and other medical fields, AI applications can improve the precision and efficiency of daily clinical practice, and can also facilitate a personalized approach to patient care, which can lead to improved outcomes and quality of care. ML has been successfully applied in various settings of daily anesthesiology practice, such as predicting acute kidney injury, optimizing anesthetic doses, and managing postoperative nausea and vomiting. The critical evaluation of ML models in healthcare is crucial to assess their validity, safety, and clinical applicability. Evaluation metrics allow an objective statistical assessment of model performance. Tools such as Shapley Values (SHAP) help interpret how individual variables contribute to model predictions. Transparency in reporting is key in maintaining trust in these technologies and to ensure their use follows ethical principles, aiming to reduce safety concerns while also benefiting patients. Understanding evaluation metrics is essential, as they provide detailed information on model performance and their ability to discriminate between individual class rates. This article offers a comprehensive framework in assessing the validity, applicability, and limitations of models, guiding responsible and effective integration of ML technologies into clinical practice. A balance between innovation, patient safety and ethical considerations must be pursued.

Machine learning approaches to injury risk prediction in sport: a scoping review with evidence synthesis

OBJECTIVE

This study reviewed the current state of machine learning (ML) research for the prediction of sports-related injuries. It aimed to chart the various approaches used and assess their efficacy, considering factors such as data heterogeneity, model specificity and contextual factors when developing predictive models.

DESIGN

Scoping review.

DATA SOURCES

PubMed, EMBASE, SportDiscus and IEEEXplore.

RESULTS

In total, 1241 studies were identified, 58 full texts were screened, and 38 relevant studies were reviewed and charted. Football (soccer) was the most commonly investigated sport. Area under the curve (AUC) was the most common means of model evaluation; it was reported in 71% of studies. In 60% of studies, tree-based solutions provided the highest statistical predictive performance. Random Forest and Extreme Gradient Boosting (XGBoost) were found to provide the highest performance for injury risk prediction. Logistic regression outperformed ML methods in 4 out of 12 studies. Three studies reported model performance of AUC>0.9, yet the clinical relevance is questionable.

CONCLUSIONS

A variety of different ML models have been applied to the prediction of sports-related injuries. While several studies report strong predictive performance, their clinical utility can be limited, with wide prediction windows or broad definitions of injury. The efficacy of ML is hampered by small datasets and numerous methodological heterogeneities (cohort sizes, definition of injury and dependent variables), which were common across the reviewed studies.

Predicting Tracheostomy Need on Admission to the Intensive Care Unit-A Multicenter Machine Learning Analysis

OBJECTIVE

It is difficult to predict which mechanically ventilated patients will ultimately require a tracheostomy which further predisposes them to unnecessary spontaneous breathing trials, additional time on the ventilator, increased costs, and further ventilation-related complications such as subglottic stenosis. In this study, we aimed to develop a machine learning tool to predict which patients need a tracheostomy at the onset of admission to the intensive care unit (ICU).

STUDY DESIGN

Retrospective Cohort Study.

SETTING

Multicenter Study of 335 Intensive Care Units between 2014 and 2015.

METHODS

The eICU Collaborative Research Database (eICU-CRD) was utilized to obtain the patient cohort. Inclusion criteria included: (1) Age >18 years and (2) ICU admission requiring mechanical ventilation. The primary outcome of interest included tracheostomy assessed via a binary classification model. Models included logistic regression (LR), random forest (RF), and Extreme Gradient Boosting (XGBoost).

RESULTS

Of 38,508 invasively mechanically ventilated patients, 1605 patients underwent a tracheostomy. The XGBoost, RF, and LR models had fair performances at an AUROC 0.794, 0.780, and 0.775 respectively. Limiting the XGBoost model to 20 features out of 331, a minimal reduction in performance was observed with an AUROC of 0.778. Using Shapley Additive Explanations, the top features were an admission diagnosis of pneumonia or sepsis and comorbidity of chronic respiratory failure.

CONCLUSIONS

Our machine learning model accurately predicts the probability that a patient will eventually require a tracheostomy upon ICU admission, and upon prospective validation, we have the potential to institute earlier interventions and reduce the complications of prolonged ventilation.

BioPrediction-RPI: Democratizing the prediction of interaction between non-coding RNA and protein with end-to-end machine learning

Machine Learning (ML) algorithms have been important tools for the extraction of useful knowledge from biological sequences, particularly in healthcare, agriculture, and the environment. However, the categorical and unstructured nature of these sequences requiring usually additional feature engineering steps, before an ML algorithm can be efficiently applied. The addition of these steps to the ML algorithm creates a processing pipeline, known as end-to-end ML. Despite the excellent results obtained by applying end-to-end ML to biotechnology problems, the performance obtained depends on the expertise of the user in the components of the pipeline. In this work, we propose an end-to-end ML-based framework called BioPrediction-RPI, which can identify implicit interactions between sequences, such as pairs of non-coding RNA and proteins, without the need for specialized expertise in end-to-end ML. This framework applies feature engineering to represent each sequence by structural and topological features. These features are divided into feature groups and used to train partial models, whose partial decisions are combined into a final decision, which, provides insights to the user by giving an interpretability report. In our experiments, the developed framework was competitive when compared with various expert-created models. We assessed BioPrediction-RPI with 12 datasets when it presented equal or better performance than all tools in 40% to 100% of cases, depending on the experiment. Finally, BioPrediction-RPI can fine-tune models based on new data and perform at the same level as ML experts, democratizing end-to-end ML and increasing its access to those working in biological sciences.

Deep learning algorithm on H&E whole slide images to characterize TP53 alterations frequency and spatial distribution in breast cancer

The tumor suppressor TP53 is frequently mutated in hormone receptor-negative, HER2-positive breast cancer (BC), contributing to tumor aggressiveness. Traditional ancillary methods like immunohistochemistry (IHC) to assess TP53 functionality face pre- and post-analytical challenges. This proof-of-concept study employed a deep learning (DL) algorithm to predict TP53 mutational status from H&E-stained whole slide images (WSIs) of BC tissue. Using a pre-trained convolutional neural network, the model identified tumor areas and predicted TP53 mutations with a Dice coefficient score of 0.82. Predictions were validated through IHC and next-generation sequencing (NGS), confirming TP53 aberrant expression in 92 % of the tumor area, closely matching IHC findings (90 %). The DL model exhibited high accuracy in tissue quantification and TP53 status prediction, outperforming traditional methods in terms of precision and efficiency. DL-based approaches offer significant promise for enhancing biomarker testing and precision oncology by reducing intra- and inter-observer variability, but further validation is required to optimize their integration into real-world clinical workflows. This study underscores the potential of DL algorithms to predict key genetic alterations, such as TP53 mutations, in BC. DL-based histopathological analysis represents a valuable tool for improving patient management and tailoring treatment approaches based on molecular biomarker status.

Survival causal rule ensemble method considering the main effect for estimating heterogeneous treatment effects

With an increasing focus on precision medicine in medical research, numerous studies have been conducted in recent years to clarify the relationship between treatment effects and patient characteristics. The treatment effects for patients with different characteristics are always heterogeneous, and therefore, various heterogeneous treatment effect machine learning estimation methods have been proposed owing to their flexibility and high estimation accuracy. However, most machine learning methods rely on black-box models, preventing direct interpretation of the relationship between patient characteristics and treatment effects. Moreover, most of these studies have focused on continuous or binary outcomes, although survival outcomes are also important in medical research. To address these challenges, we propose a heterogeneous treatment effect estimation method for survival data based on RuleFit, an interpretable machine learning method. Numerical simulation results confirmed that the prediction performance of the proposed method was comparable to that of existing methods. We also applied a dataset from an HIV study, the AIDS Clinical Trials Group Protocol 175 dataset, to illustrate the interpretability of the proposed method using real data. Consequently, the proposed survival causal rule ensemble method provides an interpretable model with sufficient estimation accuracy.

A machine-learned model for predicting weight loss success using weight change features early in treatment

Stepped-care obesity treatments aim to improve efficiency by early identification of non-responders and adjusting interventions but lack validated models. We trained a random forest classifier to improve the predictive utility of a clinical decision rule (>0.5 lb weight loss/week) that identifies non-responders in the first 2 weeks of a stepped-care weight loss trial (SMART). From 2009 to 2021, 1058 individuals with obesity participated in three studies: SMART, Opt-IN, and ENGAGED. The model was trained on 80% of the SMART data (224 participants), and its in-distribution generalizability was tested on the remaining 20% (remaining 57 participants). The out-of-distribution generalizability was tested on the ENGAGED and Opt-IN studies (472 participants). The model predicted weight loss at month 6 with an 84.5% AUROC and an 86.3% AUPRC. SHAP identified predictive features: weight loss at week 2, ranges/means and ranges of weight loss, slope, and age. The SMART-trained model showed generalizable performance with no substantial difference across studies.

Detecting irregularities in randomized controlled trials using machine learning

BACKGROUND

Over the course of a clinical trial, irregularities may arise in the data. Trialists implement human-intensive, expensive central statistical monitoring procedures to identify and correct these irregularities before the results of the trial are analyzed and disseminated. Machine learning algorithms have shown promise for identifying center-level irregularities in multi-center clinical trials with minimal human intervention. We aimed to characterize the form-level data irregularities in several historical clinical trials and evaluate the ability of a machine learning-based outlier detection algorithm to identify them.

METHODS

Data irregularities previously identified by humans in historical clinical trials were ascertained by comparing preliminary snapshots of the trial databases to the final, locked databases. We measured the ability of a machine learning based outlier detection algorithm to identify form-level irregularities using concordance (area under the receiver operator characteristic), positive predictive value (precision), and sensitivity (recall).

RESULTS

We examined preliminary snapshots of seven historical clinical trials which randomized a total of 77,001 participants. We extracted a total of 1,267,484 completed entries from 358 case report forms containing irregularities from all snapshots across all trials, containing a total of 24,850 form-wide irregularities (median per-form form-level irregularity rate: 1.81%). Our proposed machine learning algorithm detects form-level irregularities with a median concordance of 0.74 (interquartile range = 0.57-0.89), slightly exceeding the performance of a previously proposed machine learning approach with a median area under the receiver operator characteristic of 0.73 (interquartile range = 0.54-0.88).

CONCLUSION

Data irregularities in historical clinical trials were ascertained by comparing preliminary snapshots of the trial database to the final database. These irregularities can be categorized according to their scope. Irregularities can be successfully detected by a machine learning algorithm as early or earlier than a human can, without human intervention. Such an approach may complement existing techniques for central statistical monitoring in large multi-center randomized controlled trials and possibly improve the efficiency of costly data verification processes.

Construction and SHAP interpretability analysis of a risk prediction model for feeding intolerance in preterm newborns based on machine learning

OBJECTIVE

To construct a highly accurate and interpretable feeding intolerance (FI) risk prediction model for preterm newborns based on machine learning (ML) to assist medical staff in clinical diagnosis.

METHODS

In this study, a sample of 350 hospitalized preterm newborns were retrospectively analysed. First, dual feature selection was conducted to identify important feature variables for model construction. Second, ML models were constructed based on the logistic regression (LR), decision tree (DT), support vector machine (SVM) and eXtreme Gradient Boosting (XGBoost) algorithms, after which random sampling and tenfold cross-validation were separately used to evaluate and compare these models and identify the optimal model. Finally, we apply the SHapley Additive exPlanation (SHAP) interpretable framework to analyse the decision-making principles of the optimal model and expound upon the important factors affecting FI in preterm newborns and their modes of action.

RESULTS

The accuracy of XGBoost was 87.62%, and the area under the curve (AUC) was 92.2%. After the application of tenfold cross-validation, the accuracy was 83.43%, and the AUC was 89.45%, which was significantly better than those of the other models. Analysis of the XGBoost model with the SHAP interpretable framework showed that a history of resuscitation, use of probiotics, milk opening time, interval between two stools and gestational age were the main factors affecting the occurrence of FI in preterm newborns, yielding importance scores of 0.632, 0.407, 0.313, 0.313, and 0.258, respectively. A history of resuscitation, first milk opening time ≥ 24 h and interval between stools ≥ 3 days were risk factors for FI, while the use of probiotics and gestational age ≥ 34 weeks were protective factors against FI in preterm newborns.

CONCLUSIONS

In practice, we should improve perinatal care and obstetrics with the aim of reducing the occurrence of hypoxia and preterm delivery. When feeding, early milk opening, the use of probiotics, the stimulation of defecation and other measures should be implemented with the aim of reducing the occurrence of FI.

Development of a machine learning-based risk assessment model for loneliness among elderly Chinese: a cross-sectional study based on Chinese longitudinal healthy longevity survey

BACKGROUND

Loneliness is prevalent among the elderly and has intensified due to global aging trends. It adversely affects both mental and physical health. Traditional scales for measuring loneliness may yield biased results due to varying definitions. The advancements in machine learning offer new opportunities for improving the measurement and assessment of loneliness through the development of risk assessment models.

METHODS

Data from the 2018 Chinese Longitudinal Healthy Longevity Survey, involving about 16,000 participants aged ≥ 65 years, were used. The study examined the relationships between loneliness and factors such as functional limitations, living conditions, environmental influences, age-related health issues, and health behaviors. Using R 4.4.1, seven assessment models were developed: logistic regression, ridge regression, support vector machines, K-nearest neighbors, decision trees, random forests, and multi-layer perceptron. Models were evaluated based on ROC curves, accuracy, precision, recall, F1 scores, and AUC.

RESULTS

Loneliness prevalence among elderly Chinese was 23.4%. Analysis identified 15 evaluative factors and evaluated seven models. Multi-layer perceptron stands out for its strong nonlinear mapping capability and adaptability to complex data, making it one of the most effective models for assessing loneliness risk.

CONCLUSION

The study found a 23.4% prevalence of loneliness among elderly individuals in China. SHAP values indicated that marital status has the strongest evaluative value across all forecasting periods. Specifically, elderly individuals who are never married, widowed, divorced, or separated are more likely to experience loneliness compared to their married counterparts.

Application of Isokinetic Dynamometry Data in Predicting Gait Deviation Index Using Machine Learning in Stroke Patients: A Cross-Sectional Study

BACKGROUND

Three-dimensional gait analysis, supported by advanced sensor systems, is a crucial component in the rehabilitation assessment of post-stroke hemiplegic patients. However, the sensor data generated from such analyses are often complex and challenging to interpret in clinical practice, requiring significant time and complicated procedures. The Gait Deviation Index (GDI) serves as a simplified metric for quantifying the severity of pathological gait. Although isokinetic dynamometry, utilizing sophisticated sensors, is widely employed in muscle function assessment and rehabilitation, its application in gait analysis remains underexplored.

OBJECTIVE

This study aims to investigate the use of sensor-acquired isokinetic muscle strength data, combined with machine learning techniques, to predict the GDI in hemiplegic patients. This study utilizes data captured from sensors embedded in the Biodex dynamometry system and the Vicon 3D motion capture system, highlighting the integration of sensor technology in clinical gait analysis.

METHODS

This study was a cross-sectional, observational study that included a cohort of 150 post-stroke hemiplegic patients. The sensor data included measurements such as peak torque, peak torque/body weight, maximum work of repeated actions, coefficient of variation, average power, total work, acceleration time, deceleration time, range of motion, and average peak torque for both flexor and extensor muscles on the affected side at three angular velocities (60°/s, 90°/s, and 120°/s) using the Biodex System 4 Pro. The GDI was calculated using data from a Vicon 3D motion capture system. This study employed four machine learning models-Lasso Regression, Random Forest (RF), Support Vector regression (SVR), and BP Neural Network-to model and validate the sensor data. Model performance was evaluated using mean squared error (MSE), the coefficient of determination (R2), and mean absolute error (MAE). SHapley Additive exPlanations (SHAP) analysis was used to enhance model interpretability.

RESULTS

The RF model outperformed others in predicting GDI, with an MSE of 16.18, an R2 of 0.89, and an MAE of 2.99. In contrast, the Lasso Regression model yielded an MSE of 22.29, an R2 of 0.85, and an MAE of 3.71. The SVR model had an MSE of 31.58, an R2 of 0.82, and an MAE of 7.68, while the BP Neural Network model exhibited the poorest performance with an MSE of 50.38, an R2 of 0.79, and an MAE of 9.59. SHAP analysis identified the maximum work of repeated actions of the extensor muscles at 60°/s and 120°/s as the most critical sensor-derived features for predicting GDI, underscoring the importance of muscle strength metrics at varying speeds in rehabilitation assessments.

CONCLUSIONS

This study highlights the potential of integrating advanced sensor technology with machine learning techniques in the analysis of complex clinical data. The developed GDI prediction model, based on sensor-acquired isokinetic dynamometry data, offers a novel, streamlined, and effective tool for assessing rehabilitation progress in post-stroke hemiplegic patients, with promising implications for broader clinical application.

Machine Learning Identifies Arrhythmogenic Features of QRS Fragmentation in Human Cardiomyopathy: Implications for Improving Risk Stratification

BACKGROUND

Heterogeneous ventricular activation can provide the substrate for ventricular arrhythmias (VA), but its manifestation on the electrocardiogram (ECG) as a risk stratifier is not well-defined.

OBJECTIVE

To characterize the spatiotemporal features of QRS peaks that best predict VA in patients with cardiomyopathy (CM) using machine learning (ML).

METHODS

Prospectively enrolled CM patients with prophylactic defibrillators (n=95) underwent digital, high-resolution ECG recordings during intrinsic rhythm and ventricular pacing at 100 to 120 beats/min. Intra QRS peaks in the signal-averaged precordial leads were identified and their characteristics (amplitude, width, and timing within the QRS) were transformed into 4-bin histograms. Random forest models of these characteristics in each lead alongside clinical data were trained on 76 patients and tested on 19 patients with cross-validation to determine the features that predicted VA.

RESULTS

Patients were followed up for 45 (22-48) months, and 21% had VA. The individual machine learning (ML) models determined (area under the receiver operating characteristic curve [AUROC]) intrinsic QRS peak amplitude (0.88), width (0.78), and location (0.84) to all predict VA. In a combined model including all QRS peak characteristics, peaks with amplitude < 31 μV in V1, a width of 4 to 8 ms in V1, and location in the final quarter of the QRS of V1 were most predictive. Neither clinical data nor QRS peak characteristics assessed during ventricular pacing improved VA prediction when combined with intrinsic QRS peak characteristics.

CONCLUSIONS

Arrhythmogenic QRS fragmentation is characterized by narrow, low-amplitude peaks in the terminal QRS of lead V1. These features alone without clinical variables or ventricular pacing are sufficient to accurately risk stratify CM patients.

Enhancing local-scale groundwater quality predictions using advanced machine learning approaches

Assessing groundwater quality typically involves labor-intensive, time-consuming, and costly laboratory tests, making real-time monitoring impractical, especially at the local level. Groundwater quality projections at the local scale using broad spatial datasets have been inaccurate due to variations in hydrogeology, human activities, industrial operations, groundwater extraction, and waste disposal. This study aims to identify the most dependable and resilient machine learning algorithms for forecasting groundwater quality at nearby monitoring locations by utilizing simple water quality metrics that can be quickly assessed without extensive sampling and laboratory testing. The Entropy-weighted Water Quality Index (EWQI) was calculated using a large spatial and temporal dataset (2014-2021) of 977 wells with parameters including pH, total hardness (TH), calcium (Ca2⁺), magnesium (Mg2⁺), sodium (Na⁺), potassium (K⁺), sulfate (SO₄2⁻), chloride (Cl⁻), nitrate (NO₃⁻), total dissolved solids (TDS), and fluoride (F⁻). Further, similar parameters were also observed in 33 open wells at the three local monitoring sites from December 2022 to March 2023. The EWQI was predicted using a Random Forest (RF), eXtreme Gradient Boosting (XGB), and Deep Neural Network (DNN). The pH, TH, and TDS were used as input variables for EWQI predictions, as they can be easily measured using handheld probes or multi-parameters. The model performance was evaluated using R2, MAE, and RMSE. During the training stage, all three models predicted the EWQI with an R2 greater than 90%, with minimal errors when pH, TH, and TDS were input variables. To validate the models at a local scale, the EWQI was predicted at the village level (e.g., Antoli, Balapura, and Lapodiaya) using pH, TH, and TDS as input variables. The machine learning models were able to predict the EWQI very closely to the actual EWQI, with an R2 greater than 90%. It is also evident that the models could predict the EWQI using basic parameters that are easily measured, providing an overall idea of the water quality for a small area. Hence, these machine learning models could be useful for accurately representing groundwater quality, thereby avoiding the use of time-consuming and costly laboratory techniques.

Clinical usability of deep learning-based saliency maps for occlusion myocardial infarction identification from the prehospital 12-Lead electrocardiogram

INTRODUCTION

Deep learning (DL) models offer improved performance in electrocardiogram (ECG)-based classification over rule-based methods. However, for widespread adoption by clinicians, explainability methods, like saliency maps, are essential.

METHODS

On a subset of 100 ECGs from patients with chest pain, we generated saliency maps using a previously validated convolutional neural network for occlusion myocardial infarction (OMI) classification. Three clinicians reviewed ECG-saliency map dyads, first assessing the likelihood of OMI from standard ECGs and then evaluating clinical relevance and helpfulness of the saliency maps, as well as their confidence in the model's predictions. Questions were answered on a Likert scale ranging from +3 (most useful/relevant) to -3 (least useful/relevant).

RESULTS

The adjudicated accuracy of the three clinicians matched the DL model when considering area under the receiver operating characteristics curve (AUC) and F1 score (AUC 0.855 vs. 0.872, F1 score = 0.789 vs. 0.747). On average, clinicians found saliency maps slightly clinically relevant (0.96 ± 0.92) and slightly helpful (0.66 ± 0.98) in identifying or ruling out OMI but had higher confidence in the model's predictions (1.71 ± 0.56). Clinicians noted that leads I and aVL were often emphasized, even when obvious ST changes were present in other leads.

CONCLUSION

In this clinical usability study, clinicians deemed saliency maps somewhat helpful in enhancing explainability of DL-based ECG models. The spatial convolutional layers across the 12 leads in these models appear to contribute to the discrepancy between ECG segments considered most relevant by clinicians and segments that drove DL model predictions.

Tailoring Risk Prediction Models to Local Populations

Importance

Risk estimation is an integral part of cardiovascular care. Local recalibration of guideline-recommended models could address the limitations of existing tools.

Objective

To provide a machine learning (ML) approach to augment the performance of the American Heart Association's Predicting Risk of Cardiovascular Disease Events (AHA-PREVENT) equations when applied to a local population while preserving clinical interpretability.

Design, Setting, and Participants

This cohort study used a New England-based electronic health record cohort of patients without prior atherosclerotic cardiovascular disease (ASCVD) who had the data necessary to calculate the AHA-PREVENT 10-year risk of developing ASCVD in the event period (2007-2016). Patients with prior ASCVD events, death prior to 2007, or age 79 years or older in 2007 were subsequently excluded. The final study population of 95 326 patients was split into 3 nonoverlapping subsets for training, testing, and validation. The AHA-PREVENT model was adapted to this local population using the open-source ML model (MLM) Extreme Gradient Boosting model (XGBoost) with minimal predictor variables, including age, sex, and AHA-PREVENT. The MLM was monotonically constrained to preserve known associations between risk factors and ASCVD risk. Along with sex, race and ethnicity data from the electronic health record were collected to validate the performance of ASCVD risk prediction in subgroups. Data were analyzed from August 2021 to February 2024.

Main Outcomes and Measures

Consistent with the AHA-PREVENT model, ASCVD events were defined as the first occurrence of either nonfatal myocardial infarction, coronary artery disease, ischemic stroke, or cardiovascular death. Cardiovascular death was coded via government registries. Discrimination, calibration, and risk reclassification were assessed using the Harrell C index, a modified Hosmer-Lemeshow goodness-of-fit test and calibration curves, and reclassification tables, respectively.

Results

In the test set of 38 137 patients (mean [SD] age, 64.8 [6.9] years, 22 708 [59.5]% women and 15 429 [40.5%] men; 935 [2.5%] Asian, 2153 [5.6%] Black, 1414 [3.7%] Hispanic, 31 400 [82.3%] White, and 2235 [5.9%] other, including American Indian, multiple races, unspecified, and unrecorded, consolidated owing to small numbers), MLM-PREVENT had improved calibration (modified Hosmer-Lemeshow P > .05) compared to the AHA-PREVENT model across risk categories in the overall cohort (χ23 = 2.2; P = .53 vs χ23 > 16.3; P < .001) and sex subgroups (men: χ23 = 2.1; P = .55 vs χ23 > 16.3; P < .001; women: χ23 = 6.5; P = .09 vs. χ23 > 16.3; P < .001), while also surpassing a traditional recalibration approach. MLM-PREVENT maintained or improved AHA-PREVENT's calibration in Asian, Black, and White individuals. Both MLM-PREVENT and AHA-PREVENT performed equally well in discriminating risk (approximate ΔC index, ±0.01). Using a clinically significant 7.5% risk threshold, MLM-PREVENT reclassified a total of 11.5% of patients. We visualize the recalibration through MLM-PREVENT ASCVD risk charts that highlight preserved risk associations of the original AHA-PREVENT model.

Conclusions and Relevance

The interpretable ML approach presented in this article enhanced the accuracy of the AHA-PREVENT model when applied to a local population while still preserving the risk associations found by the original model. This method has the potential to recalibrate other established risk tools and is implementable in electronic health record systems for improved cardiovascular risk assessment.

Essentials for AI Research in Cardiology: Challenges and Mitigations

Technology using artificial intelligence (AI) is flourishing; the same advancements can be seen in health care. Cardiology in particular is well placed to take advantage of AI because of the data-intensive nature of the field and the current strain on existing resources in the management of cardiovascular disease. With AI nearing the stage of routine implementation into clinical care, considerations need to be made to ensure the software is effective and safe. The benefits of AI are well established, but the challenges and ethical considerations are less well understood. As a result, there is currently a lack of consensus on what the essential components are in an AI study. In this review we aim to assess and provide greater clarity on the challenges encountered in conducting AI studies and explore potential mitigations that could facilitate the successful integration of AI in the management of cardiovascular disease.

CardioMEA: comprehensive data analysis platform for studying cardiac diseases and drug responses

Introduction

In recent years, high-density microelectrode arrays (HD-MEAs) have emerged as a valuable tool in preclinical research for characterizing the electrophysiology of human induced pluripotent stem-cell-derived cardiomyocytes (iPSC-CMs). HD-MEAs enable the capturing of both extracellular and intracellular signals on a large scale, while minimizing potential damage to the cell. However, despite technological advancements of HD-MEAs, there is a lack of effective data-analysis platforms that are capable of processing and analyzing the data, particularly in the context of cardiac arrhythmias and drug testing.

Methods

To address this need, we introduce CardioMEA, a comprehensive data-analysis platform designed specifically for HD-MEA data that have been obtained from iPSCCMs. CardioMEA features scalable data processing pipelines and an interactive web-based dashboard for advanced visualization and analysis. In addition to its core functionalities, CardioMEA incorporates modules designed to discern crucial electrophysiological features between diseased and healthy iPSC-CMs. Notably, CardioMEA has the unique capability to analyze both extracellular and intracellular signals, thereby facilitating customized analyses for specific research tasks.

Results and discussion

We demonstrate the practical application of CardioMEA by analyzing electrophysiological signals from iPSC-CM cultures exposed to seven antiarrhythmic drugs. CardioMEA holds great potential as an intuitive, userfriendly platform for studying cardiac diseases and assessing drug effects.

Implementation of Machine Learning Algorithms to Screen for Advanced Liver Fibrosis in Metabolic Dysfunction-Associated Steatotic Liver Disease: An In-Depth Explanatory Analysis

INTRODUCTION

This study aimed to train machine learning algorithms (MLAs) to detect advanced fibrosis (AF) in metabolic dysfunction-associated steatotic liver disease (MASLD) patients at the level of primary care setting and to explain the predictions to ensure responsible use by clinicians.

METHODS

Readily available features of 618 MASLD patients followed up at a tertiary center were used to train five MLAs. AF was defined as liver stiffness ≥9.3 kPa, measured via 2-dimension shear wave elastography (n = 495) or liver biopsy ≥F3 (n = 123). MLAs were compared to Fibrosis-4 index (FIB-4) and non-alcoholic fatty liver disease (NAFLD) fibrosis score (NFS) on 540 MASLD patients from the primary care setting as validation. Feature importance, partial dependence, and shapely additive explanations (SHAPs) were utilized for explanation.

RESULTS

Extreme gradient boosting (XGBoost) achieved an AUC = 0.91, outperforming FIB-4 (AUC = 0.78) and NFS (AUC = 0.81, both p < 0.05) with specificity = 76% versus 59% and 48% for FIB-4 ≥1.3 and NFS ≥-1.45, respectively (p < 0.05). Its sensitivity (91%) was superior to FIB-4 (79%). XGBoost confidently excluded AF (negative predictive value = 99%) with the highest positive predictive value (31%), superior to FIB-4 and NFS (all p < 0.05). The most important features were HbA1c and gamma glutamyl transpeptidase (GGT) with a steep increase in AF probability at HbA1c >6.5%. The strongest interaction was between AST and age. XGBoost, but not logistic regression, extracted informative patterns from ALT, low-density lipoprotein cholesterol, and alkaline phosphatase (p < 0.001). One-quarter of the false positives (FPs) were correctly reclassified with only one additional false negative based on the SHAP values of GGT, platelets, and ALT which were found to be associated with a FP classification.

CONCLUSIONS

An explainable XGBoost algorithm was demonstrated superior to FIB-4 and NFS for screening of AF in MASLD patients at the primary care setting. The algorithm also proved safe for use as clinicians can understand the predictions and flag FP classifications.

Enhancing transparency and fairness in automated credit decisions: an explainable novel hybrid machine learning approach

This paper uses a generalised stacking method to introduce a novel hybrid model that combines a one-dimensional convolutional neural network 1DCNN with extreme gradient boosting XGBoost. We compared the predictive accuracies of the proposed hybrid architecture with three conventional algorithms-1DCNN, XGBoost and logistic regression (LR) using a dataset of over twenty thousand peer-to-peer (P2P) consumer credit observations. By leveraging the SHAP algorithm, the research provides a detailed analysis of feature importance, contributing to the model's predictions and offering insights into the overall and individual significance of different features. The findings demonstrate that the hybrid model outperforms the LR, XGBoost and 1DCNN models in terms of classification accuracy. Furthermore, the research addresses concern regarding fairness and bias by showing that removing potentially discriminatory features, such as age and gender, does not significantly impact the hybrid model's classification capabilities. This suggests that fair and unbiased credit scoring models can achieve high effectiveness levels without compromising accuracy. This paper makes significant contributions to academic research and practical applications in credit risk management by presenting a hybrid model that offers superior classification accuracy and promotes interpretability using the model agnostic SHAP framework.

Anatomic Interpretability in Neuroimage Deep Learning: Saliency Approaches for Typical Aging and Traumatic Brain Injury

The black box nature of deep neural networks (DNNs) makes researchers and clinicians hesitant to rely on their findings. Saliency maps can enhance DNN explainability by suggesting the anatomic localization of relevant brain features. This study compares seven popular attribution-based saliency approaches to assign neuroanatomic interpretability to DNNs that estimate biological brain age (BA) from magnetic resonance imaging (MRI). Cognitively normal (CN) adults ( N = 13,394 , 5,900 males; mean age: 65.82 ± 8.89 years) are included for DNN training, testing, validation, and saliency map generation to estimate BA. To study saliency robustness to the presence of anatomic deviations from normality, saliency maps are also generated for adults with mild traumatic brain injury (mTBI, N = 214 , 135 males; mean age: 55.3 ± 9.9 years). We assess saliency methods' capacities to capture known anatomic features of brain aging and compare them to a surrogate ground truth whose anatomic saliency is known a priori. Anatomic aging features are identified most reliably by the integrated gradients method, which outperforms all others through its ability to localize relevant anatomic features. Gradient Shapley additive explanations, input × gradient, and masked gradient perform less consistently but still highlight ubiquitous neuroanatomic features of aging (ventricle dilation, hippocampal atrophy, sulcal widening). Saliency methods involving gradient saliency, guided backpropagation, and guided gradient-weight class attribution mapping localize saliency outside the brain, which is undesirable. Our research suggests the relative tradeoffs of saliency methods to interpret DNN findings during BA estimation in typical aging and after mTBI.

Interpretable Clinical Decision-Making Application for Etiological Diagnosis of Ventricular Tachycardia Based on Machine Learning

Background: Ventricular tachycardia (VT) can broadly be categorised into ischemic heart disease, non-ischemic structural heart disease, and idiopathic VT. There are few studies related to the application of machine learning for the etiological diagnosis of VT, and the interpretable methods are still in the exploratory stage for clinical decision-making applications. Objectives: The aim is to propose a machine learning model for the etiological diagnosis of VT. Interpretable results based on models are compared with expert knowledge, and interpretable evaluation protocols for clinical decision-making applications are developed. Methods: A total of 1305 VT patient data from 1 January 2013 to 1 September 2023 at the Arrhythmia Centre of Fuwai Hospital were included in the study. Clinical data collected during hospitalisation included demographics, medical history, vital signs, echocardiographic results, and laboratory test outcomes. Results: The XGBoost model demonstrated the best performance in VT etiological diagnosis (precision, recall, and F1 were 88.4%, 88.5%, and 88.4%, respectively). A total of four interpretable machine learning methods applicable to clinical decision-making were evaluated in terms of visualisation, clinical usability, clinical applicability, and efficiency with expert knowledge interpretation. Conclusions: The XGBoost model demonstrated superior performance in the etiological diagnosis of VT, and SHAP and decision tree interpretable methods are more favoured by clinicians for decision-making.

Digital health in cardiovascular medicine: An overview of key applications and clinical impact by the Portuguese Society of Cardiology Study Group on Digital Health

Digital health interventions including telehealth, mobile health, artificial intelligence, big data, robotics, extended reality, computational and high-fidelity bench simulations are an integral part of the path toward precision medicine. Current applications encompass risk factor modification, chronic disease management, clinical decision support, diagnostics interpretation, preprocedural planning, evidence generation, education, and training. Despite the acknowledged potential, their development and implementation have faced several challenges and constraints, meaning few digital health tools have reached daily clinical practice. As a result, the Portuguese Society of Cardiology Study Group on Digital Health set out to outline the main digital health applications, address some of the roadblocks hampering large-scale deployment, and discuss future directions in support of cardiovascular health at large.

Prediction of deoxynivalenol contamination in spring oats in Sweden using explainable artificial intelligence

Weather conditions and agronomical factors are known to affect Fusarium spp. growth and ultimately deoxynivalenol (DON) contamination in oat. This study aimed to develop predictive models for the contamination of spring oat at harvest with DON on a regional basis in Sweden using machine-learning algorithms. Three models were developed as regional risk-assessment tools for farmers, crop collectors, and food safety inspectors, respectively. Data included: weather data from different oat growing periods, agronomical data, site-specific data, and DON contamination data from the previous year. Results showed that: (1) RF models were able to predict DON contamination at harvest with a total classification accuracy of minimal 0.72; (2) good predictions could already be made in June; (3) rainfall, relative humidity, and wind speed in different oat growing stages, followed by crop variety and elevation were the most important features for predicting DON contamination in spring oats at harvest.

Prediction models for identifying medication overuse or medication overuse headache in migraine patients: a systematic review

BACKGROUND

Migraine is a debilitating neurological disorder that presents significant management challenges, resulting in underdiagnosis and inappropriate treatments, leaving patients at risk of medication overuse (MO). MO contributes to disease progression and the development of medication overuse headache (MOH). Predicting which migraine patients are at risk of MO/MOH is crucial for effective management. Thus, this systematic review aims to review and critique available prediction models for MO/MOH in migraine patients.

METHODS

A systematic search was conducted using Embase, Scopus, Medline/PubMed, ACM Digital Library, and IEEE databases from inception to April 22, 2024. The risk of bias was assessed using the prediction model risk of bias assessment tool.

RESULTS

Out of 1,579 articles, six studies with nine models met the inclusion criteria. Three studies developed new prediction models, while the remaining validated existing scores. Most studies utilized cross-sectional and prospective data collection in specific headache settings and migraine types. The models included up to 53 predictors, with sample sizes from 17 to 1,419 participants. Traditional statistical models (logistic regression and least absolute shrinkage and selection operator regression) were used in two studies, while one utilized a machine learning (ML) technique (support vector machines). Receiver operating characteristic analysis was employed to validate existing scores. The area under the receiver operating characteristic (AUROC) for the ML model (0.83) outperformed the traditional statistical model (0.62) in internal validation. The AUROCs ranged from 0.84 to 0.85 for the validation of existing scores. Common predictors included age and gender; genetic data and questionnaire evaluations were also included. All studies demonstrated a high risk of bias in model construction and high concerns regarding applicability to participants.

CONCLUSION

This review identified promising results for MO/MOH prediction models in migraine patients, although the field remains limited. Future research should incorporate important risk factors, assess discrimination and calibration, and perform external validation. Further studies with robust designs, appropriate settings, high-quality and quantity data, and rigorous methodologies are necessary to advance this field.

Unlocking Tomorrow's Health Care: Expanding the Clinical Scope of Wearables by Applying Artificial Intelligence

As an integral aspect of health care, digital technology has enabled modelling of complex relationships to detect, screen, diagnose, and predict patient outcomes. With massive data sets, artificial intelligence (AI) can have marked effects on 3 levels: for patients, clinicians, and health systems. In this review, we discuss contemporary AI-enabled wearable devices undergoing research in the field of cardiovascular medicine. These include devices such as smart watches, electrocardiogram patches, and smart textiles such as smart socks and chest sensors for diagnosis, management, and prognostication of conditions such as atrial fibrillation, heart failure, and hypertension as well as monitoring for cardiac rehabilitation. We review the evolution of machine learning algorithms used in wearable devices from random forest models to the use of convolutional neural networks and transformers. We further discuss frameworks for wearable technologies such as the V3-stage process of verification, analytical validation, and clinical validation as well as challenges of AI integration in medicine such as data veracity, validity, and security and provide a reference framework to maintain fairness and equity. Last, clinician and patient perspectives are discussed to highlight the importance of considering end-user feedback in development and regulatory processes.

Anatomic Interpretability in Neuroimage Deep Learning: Saliency Approaches for Typical Aging and Traumatic Brain Injury

The black box nature of deep neural networks (DNNs) makes researchers and clinicians hesitant to rely on their findings. Saliency maps can enhance DNN explainability by suggesting the anatomic localization of relevant brain features. This study compares seven popular attribution-based saliency approaches to assign neuroanatomic interpretability to DNNs that estimate biological brain age (BA) from magnetic resonance imaging (MRI). Cognitively normal (CN) adults (N = 13,394, 5,900 males; mean age: 65.82 ± 8.89 years) are included for DNN training, testing, validation, and saliency map generation to estimate BA. To study saliency robustness to the presence of anatomic deviations from normality, saliency maps are also generated for adults with mild traumatic brain injury (mTBI, N = 214, 135 males; mean age: 55.3 ± 9.9 years). We assess saliency methods' capacities to capture known anatomic features of brain aging and compare them to a surrogate ground truth whose anatomic saliency is known a priori. Anatomic aging features are identified most reliably by the integrated gradients method, which outperforms all others through its ability to localize relevant anatomic features. Gradient Shapley additive explanations, input × gradient, and masked gradient perform less consistently but still highlight ubiquitous neuroanatomic features of aging (ventricle dilation, hippocampal atrophy, sulcal widening). Saliency methods involving gradient saliency, guided backpropagation, and guided gradient-weight class attribution mapping localize saliency outside the brain, which is undesirable. Our research suggests the relative tradeoffs of saliency methods to interpret DNN findings during BA estimation in typical aging and after mTBI.

Potentials of artificial intelligence in familial hypercholesterolemia: Advances in screening, diagnosis, and risk stratification for early intervention and treatment

Familial hypercholesterolemia (FH) poses a global health challenge due to high incidence rates and underdiagnosis, leading to increased risks of early-onset atherosclerosis and cardiovascular diseases. Early detection and treatment of FH is critical in reducing the risk of cardiovascular events and improving the long-term outcomes and quality of life for affected individuals and their families. Traditional therapeutic approaches revolve around lipid-lowering interventions, yet challenges persist, particularly in accurate and timely diagnosis. The current diagnostic landscape heavily relies on genetic testing of specific LDL-C metabolism genes, often limited to specialized centers. This constraint has led to the adoption of alternative clinical scores for FH diagnosis. However, the rapid advancements in artificial intelligence (AI) and machine learning (ML) present promising solutions to these diagnostic challenges. This review explores the intricacies of FH, highlighting the challenges that are encountered in the diagnosis and management of the disorder. The revolutionary potential of ML, particularly in large-scale population screening, is highlighted. Applications of ML in FH screening, diagnosis, and risk stratification are discussed, showcasing its ability to outperform traditional criteria. However, challenges and ethical considerations, including algorithmic stability, data quality, privacy, and consent issues, are crucial areas that require attention. The review concludes by emphasizing the significant promise of AI and ML in FH management while underscoring the need for ethical and practical vigilance to ensure responsible and effective integration into healthcare practices.

Personalized Prediction of Long-Term Renal Function Prognosis Following Nephrectomy Using Interpretable Machine Learning Algorithms: Case-Control Study

BACKGROUND

Acute kidney injury (AKI) is a common adverse outcome following nephrectomy. The progression from AKI to acute kidney disease (AKD) and subsequently to chronic kidney disease (CKD) remains a concern; yet, the predictive mechanisms for these transitions are not fully understood. Interpretable machine learning (ML) models offer insights into how clinical features influence long-term renal function outcomes after nephrectomy, providing a more precise framework for identifying patients at risk and supporting improved clinical decision-making processes.

OBJECTIVE

This study aimed to (1) evaluate postnephrectomy rates of AKI, AKD, and CKD, analyzing long-term renal outcomes along different trajectories; (2) interpret AKD and CKD models using Shapley Additive Explanations values and Local Interpretable Model-Agnostic Explanations algorithm; and (3) develop a web-based tool for estimating AKD or CKD risk after nephrectomy.

METHODS

We conducted a retrospective cohort study involving patients who underwent nephrectomy between July 2012 and June 2019. Patient data were randomly split into training, validation, and test sets, maintaining a ratio of 76.5:8.5:15. Eight ML algorithms were used to construct predictive models for postoperative AKD and CKD. The performance of the best-performing models was assessed using various metrics. We used various Shapley Additive Explanations plots and Local Interpretable Model-Agnostic Explanations bar plots to interpret the model and generated directed acyclic graphs to explore the potential causal relationships between features. Additionally, we developed a web-based prediction tool using the top 10 features for AKD prediction and the top 5 features for CKD prediction.

RESULTS

The study cohort comprised 1559 patients. Incidence rates for AKI, AKD, and CKD were 21.7% (n=330), 15.3% (n=238), and 10.6% (n=165), respectively. Among the evaluated ML models, the Light Gradient-Boosting Machine (LightGBM) model demonstrated superior performance, with an area under the receiver operating characteristic curve of 0.97 for AKD prediction and 0.96 for CKD prediction. Performance metrics and plots highlighted the model's competence in discrimination, calibration, and clinical applicability. Operative duration, hemoglobin, blood loss, urine protein, and hematocrit were identified as the top 5 features associated with predicted AKD. Baseline estimated glomerular filtration rate, pathology, trajectories of renal function, age, and total bilirubin were the top 5 features associated with predicted CKD. Additionally, we developed a web application using the LightGBM model to estimate AKD and CKD risks.

CONCLUSIONS

An interpretable ML model effectively elucidated its decision-making process in identifying patients at risk of AKD and CKD following nephrectomy by enumerating critical features. The web-based calculator, found on the LightGBM model, can assist in formulating more personalized and evidence-based clinical strategies.

Ensembling U-Nets for microaneurysm segmentation in optical coherence tomography angiography in patients with diabetic retinopathy

Diabetic retinopathy is one of the leading causes of blindness around the world. This makes early diagnosis and treatment important in preventing vision loss in a large number of patients. Microaneurysms are the key hallmark of the early stage of the disease, non-proliferative diabetic retinopathy, and can be detected using OCT angiography quickly and non-invasively. Screening tools for non-proliferative diabetic retinopathy using OCT angiography thus have the potential to lead to improved outcomes in patients. We compared different configurations of ensembled U-nets to automatically segment microaneurysms from OCT angiography fundus projections. For this purpose, we created a new database to train and evaluate the U-nets, created by two expert graders in two stages of grading. We present the first U-net neural networks using ensembling for the detection of microaneurysms from OCT angiography en face images from the superficial and deep capillary plexuses in patients with non-proliferative diabetic retinopathy trained on a database labeled by two experts with repeats.

Can Loneliness be Predicted? Development of a Risk Prediction Model for Loneliness among Elderly Chinese: A Study Based on CLHLS

Background

Loneliness is prevalent among the elderly, worsened by global aging trends. It impacts mental and physiological health. Traditional scales for measuring loneliness may be biased due to cognitive decline and varying definitions. Machine learning advancements offer potential improvements in risk prediction models.

Methods

Data from the 2018 Chinese Longitudinal Healthy Longevity Survey (CLHLS), involving over 16,000 participants aged ≥65 years, were used. The study examined the relationships between loneliness and factors such as cognitive function, functional limitations, living conditions, environmental influences, age-related health issues, and health behaviors. Using R 4.4.1, seven predictive models were developed: logistic regression, ridge regression, support vector machines, K-nearest neighbors, decision trees, random forests, and multi-layer perceptron. Models were evaluated based on ROC curves, accuracy, precision, recall, F1 scores, and AUC.

Results

Loneliness prevalence among elderly Chinese was 23.4%. Analysis identified 16 predictive factors and evaluated seven models. Logistic regression was the most effective model for predicting loneliness risk due to its economic and operational advantages.

Conclusion

The study found a 23.4% prevalence of loneliness among elderly individuals in China. SHAP values indicated that higher MMSE scores correlate with lower loneliness levels. Logistic regression was the superior model for predicting loneliness risk in this population.

Intelligent systems in healthcare: A systematic survey of explainable user interfaces

With radiology shortages affecting over half of the global population, the potential of artificial intelligence to revolutionize medical diagnosis and treatment is ever more important. However, lacking trust from medical professionals hinders the widespread adoption of AI models in health sciences. Explainable AI (XAI) aims to increase trust and understanding of black box models by identifying biases and providing transparent explanations. This is the first survey that explores explainable user interfaces (XUI) from a medical domain perspective, analysing the visualization and interaction methods employed in current medical XAI systems. We analysed 42 explainable interfaces following the PRISMA methodology, emphasizing the critical role of effectively conveying information to users as part of the explanation process. We contribute a taxonomy of interface design properties and identify five distinct clusters of research papers. Future research directions include contestability in medical decision support, counterfactual explanations for images, and leveraging Large Language Models to enhance XAI interfaces in healthcare.

Impact of high-quality, mixed-domain data on the performance of medical language models

OBJECTIVE

To optimize the training strategy of large language models for medical applications, focusing on creating clinically relevant systems that efficiently integrate into healthcare settings, while ensuring high standards of accuracy and reliability.

MATERIALS AND METHODS

We curated a comprehensive collection of high-quality, domain-specific data and used it to train several models, each with different subsets of this data. These models were rigorously evaluated against standard medical benchmarks, such as the USMLE, to measure their performance. Furthermore, for a thorough effectiveness assessment, they were compared with other state-of-the-art medical models of comparable size.

RESULTS

The models trained with a mix of high-quality, domain-specific, and general data showed superior performance over those trained on larger, less clinically relevant datasets (P < .001). Our 7-billion-parameter model Med5 scores 60.5% on MedQA, outperforming the previous best of 49.3% from comparable models, and becomes the first of its size to achieve a passing score on the USMLE. Additionally, this model retained its proficiency in general domain tasks, comparable to state-of-the-art general domain models of similar size.

DISCUSSION

Our findings underscore the importance of integrating high-quality, domain-specific data in training large language models for medical purposes. The balanced approach between specialized and general data significantly enhances the model's clinical relevance and performance.

CONCLUSION

This study sets a new standard in medical language models, proving that a strategically trained, smaller model can outperform larger ones in clinical relevance and general proficiency, highlighting the importance of data quality and expert curation in generative artificial intelligence for healthcare applications.

Comparing code-free and bespoke deep learning approaches in ophthalmology

AIM

Code-free deep learning (CFDL) allows clinicians without coding expertise to build high-quality artificial intelligence (AI) models without writing code. In this review, we comprehensively review the advantages that CFDL offers over bespoke expert-designed deep learning (DL). As exemplars, we use the following tasks: (1) diabetic retinopathy screening, (2) retinal multi-disease classification, (3) surgical video classification, (4) oculomics and (5) resource management.

METHODS

We performed a search for studies reporting CFDL applications in ophthalmology in MEDLINE (through PubMed) from inception to June 25, 2023, using the keywords 'autoML' AND 'ophthalmology'. After identifying 5 CFDL studies looking at our target tasks, we performed a subsequent search to find corresponding bespoke DL studies focused on the same tasks. Only English-written articles with full text available were included. Reviews, editorials, protocols and case reports or case series were excluded. We identified ten relevant studies for this review.

RESULTS

Overall, studies were optimistic towards CFDL's advantages over bespoke DL in the five ophthalmological tasks. However, much of such discussions were identified to be mono-dimensional and had wide applicability gaps. High-quality assessment of better CFDL applicability over bespoke DL warrants a context-specific, weighted assessment of clinician intent, patient acceptance and cost-effectiveness. We conclude that CFDL and bespoke DL are unique in their own assets and are irreplaceable with each other. Their benefits are differentially valued on a case-to-case basis. Future studies are warranted to perform a multidimensional analysis of both techniques and to improve limitations of suboptimal dataset quality, poor applicability implications and non-regulated study designs.

CONCLUSION

For clinicians without DL expertise and easy access to AI experts, CFDL allows the prototyping of novel clinical AI systems. CFDL models concert with bespoke models, depending on the task at hand. A multidimensional, weighted evaluation of the factors involved in the implementation of those models for a designated task is warranted.

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.

Exploring explainable AI features in the vocal biomarkers of lung disease

This review delves into the burgeoning field of explainable artificial intelligence (XAI) in the detection and analysis of lung diseases through vocal biomarkers. Lung diseases, often elusive in their early stages, pose a significant public health challenge. Recent advancements in AI have ushered in innovative methods for early detection, yet the black-box nature of many AI models limits their clinical applicability. XAI emerges as a pivotal tool, enhancing transparency and interpretability in AI-driven diagnostics. This review synthesizes current research on the application of XAI in analyzing vocal biomarkers for lung diseases, highlighting how these techniques elucidate the connections between specific vocal features and lung pathology. We critically examine the methodologies employed, the types of lung diseases studied, and the performance of various XAI models. The potential for XAI to aid in early detection, monitor disease progression, and personalize treatment strategies in pulmonary medicine is emphasized. Furthermore, this review identifies current challenges, including data heterogeneity and model generalizability, and proposes future directions for research. By offering a comprehensive analysis of explainable AI features in the context of lung disease detection, this review aims to bridge the gap between advanced computational approaches and clinical practice, paving the way for more transparent, reliable, and effective diagnostic tools.

Development and external validation of temporal fusion transformer models for continuous intraoperative blood pressure forecasting

Background

During surgery, intraoperative hypotension is associated with postoperative morbidity and should therefore be avoided. Predicting the occurrence of hypotension in advance may allow timely interventions to prevent hypotension. Previous prediction models mostly use high-resolution waveform data, which is often not available.

Methods

We utilised a novel temporal fusion transformer (TFT) algorithm to predict intraoperative blood pressure trajectories 7 min in advance. We trained the model with low-resolution data (sampled every 15 s) from 73,009 patients who were undergoing general anaesthesia for non-cardiothoracic surgery between January 1, 2017, and December 30, 2020, at the General Hospital of Vienna, Austria. The data set contained information on patient demographics, vital signs, medication, and ventilation. The model was evaluated using an internal (n = 8113) and external test set (n = 5065) obtained from the openly accessible Vital Signs Database.

Findings

In the internal test set, the mean absolute error for predicting mean arterial blood pressure was 0.376 standard deviations-or 4 mmHg-and 0.622 standard deviations-or 7 mmHg-in the external test set. We also adapted the TFT model to binarily predict the occurrence of hypotension as defined by mean arterial blood pressure < 65 mmHg in the next one, three, five, and 7 min. Here, model discrimination was excellent, with a mean area under the receiver operating characteristic curve (AUROC) of 0.933 in the internal test set and 0.919 in the external test set.

Interpretation

Our TFT model is capable of accurately forecasting intraoperative arterial blood pressure using only low-resolution data showing a low prediction error. When used for binary prediction of hypotension, we obtained excellent performance.

Funding

No external funding.

Interpretable machine learning for dermatological disease detection: Bridging the gap between accuracy and explainability

Research on disease detection by leveraging machine learning techniques has been under significant focus. The use of machine learning techniques is important to detect critical diseases promptly and provide the appropriate treatment. Disease detection is a vital and sensitive task and while machine learning models may provide a robust solution, they can come across as complex and unintuitive. Therefore, it is important to gauge a better understanding of the predictions and trust the results. This paper takes up the crucial task of skin disease detection and introduces a hybrid machine learning model combining SVM and XGBoost for the detection task. The proposed model outperformed the existing machine learning models - Support Vector Machine (SVM), decision tree, and XGBoost with an accuracy of 99.26%. The increased accuracy is essential for detecting skin disease due to the similarity in the symptoms which make it challenging to differentiate between the different conditions. In order to foster trust and gain insights into the results we turn to the promising field of Explainable Artificial Intelligence (XAI). We explore two such frameworks for local as well as global explanations for these machine learning models namely, SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME).

Explainable and visualizable machine learning models to predict biochemical recurrence of prostate cancer

PURPOSE

Machine learning (ML) models presented an excellent performance in the prognosis prediction. However, the black box characteristic of ML models limited the clinical applications. Here, we aimed to establish explainable and visualizable ML models to predict biochemical recurrence (BCR) of prostate cancer (PCa).

MATERIALS AND METHODS

A total of 647 PCa patients were retrospectively evaluated. Clinical parameters were identified using LASSO regression. Then, cohort was split into training and validation datasets with a ratio of 0.75:0.25 and BCR-related features were included in Cox regression and five ML algorithm to construct BCR prediction models. The clinical utility of each model was evaluated by concordance index (C-index) values and decision curve analyses (DCA). Besides, Shapley Additive Explanation (SHAP) values were used to explain the features in the models.

RESULTS

We identified 11 BCR-related features using LASSO regression, then establishing five ML-based models, including random survival forest (RSF), survival support vector machine (SSVM), survival Tree (sTree), gradient boosting decision tree (GBDT), extreme gradient boosting (XGBoost), and a Cox regression model, C-index were 0.846 (95%CI 0.796-0.894), 0.774 (95%CI 0.712-0.834), 0.757 (95%CI 0.694-0.818), 0.820 (95%CI 0.765-0.869), 0.793 (95%CI 0.735-0.852), and 0.807 (95%CI 0.753-0.858), respectively. The DCA showed that RSF model had significant advantages over all models. In interpretability of ML models, the SHAP value demonstrated the tangible contribution of each feature in RSF model.

CONCLUSIONS

Our score system provide reference for the identification for BCR, and the crafting of a framework for making therapeutic decisions for PCa on a personalized basis.

Artificial intelligence in pancreaticobiliary endoscopy: Current applications and future directions

Pancreaticobiliary endoscopy is an essential tool for diagnosing and treating pancreaticobiliary diseases. However, it does not fully meet clinical needs, which presents challenges such as significant difficulty in operation and risks of missed diagnosis or misdiagnosis. In recent years, artificial intelligence (AI) has enhanced the diagnostic and treatment efficiency and quality of pancreaticobiliary endoscopy. Diagnosis and differential diagnosis based on endoscopic ultrasound (EUS) images, pathology of EUS-guided fine-needle aspiration or biopsy, need for endoscopic retrograde cholangiopancreatography (ERCP) and assessment of operational difficulty, postoperative complications and prediction of patient prognosis, and real-time procedure guidance. This review provides an overview of AI applications in pancreaticobiliary endoscopy and proposes future development directions in aspects such as data quality and algorithmic interpretability, aiming to provide new insights for the integration of AI technology with pancreaticobiliary endoscopy.

The impact of AI-enabled CRM systems on organizational competitive advantage: A mixed-method approach using BERTopic and PLS-SEM

The recent advances in machine learning and deep learning algorithms, along with the advent of generative AI, have led AI to become the "new normal" in organizations. This trend has extended to CRM, resulting in the development of AI-enabled CRM systems, or AI-CRM. Despite the growing adoption of AI as part of competitive strategies, many firms report minimal or no positive effect of AI on performance. This study addresses the research questions: "What are the critical features of AI-CRM systems?" and "How do these features impact organizational competitive advantage?" To explore this, we aim to identify key characteristics of AI-CRM and assess their impact on organizational performance. In Study 1, we utilize BERTopic topic modeling to extract critical features of AI-CRM from user reviews. Study 2 employs PLS-SEM to examine how these features influence organizational competitive advantage. Study 1 reveals four main characteristics of AI-CRM (general, marketing, sales, and service/support), each comprising distinct features. Study 2 shows that these characteristics differentially impact CRM capability, significantly affecting performance and competitive advantage. The findings offer valuable insights for both theory and practice regarding the effective use of AI in organizations.

Using machine learning methods to predict all-cause somatic hospitalizations in adults: A systematic review

AIM

In this review, we investigated how Machine Learning (ML) was utilized to predict all-cause somatic hospital admissions and readmissions in adults.

METHODS

We searched eight databases (PubMed, Embase, Web of Science, CINAHL, ProQuest, OpenGrey, WorldCat, and MedNar) from their inception date to October 2023, and included records that predicted all-cause somatic hospital admissions and readmissions of adults using ML methodology. We used the CHARMS checklist for data extraction, PROBAST for bias and applicability assessment, and TRIPOD for reporting quality.

RESULTS

We screened 7,543 studies of which 163 full-text records were read and 116 met the review inclusion criteria. Among these, 45 predicted admission, 70 predicted readmission, and one study predicted both. There was a substantial variety in the types of datasets, algorithms, features, data preprocessing steps, evaluation, and validation methods. The most used types of features were demographics, diagnoses, vital signs, and laboratory tests. Area Under the ROC curve (AUC) was the most used evaluation metric. Models trained using boosting tree-based algorithms often performed better compared to others. ML algorithms commonly outperformed traditional regression techniques. Sixteen studies used Natural language processing (NLP) of clinical notes for prediction, all studies yielded good results. The overall adherence to reporting quality was poor in the review studies. Only five percent of models were implemented in clinical practice. The most frequently inadequately addressed methodological aspects were: providing model interpretations on the individual patient level, full code availability, performing external validation, calibrating models, and handling class imbalance.

CONCLUSION

This review has identified considerable concerns regarding methodological issues and reporting quality in studies investigating ML to predict hospitalizations. To ensure the acceptability of these models in clinical settings, it is crucial to improve the quality of future studies.

An interpretable ensemble trees method with joint analysis of static and dynamic features for myocardial infarction detection

Objective.In recent years, artificial intelligence-based electrocardiogram (ECG) methods have been massively applied to myocardial infarction (MI). However, the joint analysis of static and dynamic features to achieve accurate and interpretable MI detection has not been comprehensively addressed.Approach.This paper proposes a simplified ensemble tree method with a joint analysis of static and dynamic features to solve this issue for MI detection. Initially, the dynamic features are extracted by modeling the intrinsic dynamics of ECG via dynamic learning in addition to extracting classical static features. Secondly, a two-stage feature selection strategy is designed to identify a few significant features, which substitute the original variables that are employed in constructing the ensemble tree. This approach enhances the discriminative ability by selecting significant static and dynamic features. Subsequently, this paper presents an interpretable classification method named StackTree by introducing a stacked ensemble scheme to modify the ensemble tree simplification algorithm. The representative rules of the raw ensemble trees are selected as the intermediate training data that is used to retrain a decision tree with performance close to that of the source ensemble model. Using this scheme, the significant precision and interpretability of MI detection are thus comprehensively addressed.Main results.The effectiveness of our method in detecting MI is evaluated using the Physikalisch-Technische Bundesanstalt (PTB) and clinical database. The findings suggest that our algorithm outperforms the traditional methods based on a single type of feature. Additionally, it is comparable to the conventional random forest, achieving 97.1% accuracy under the inter-patient framework on the PTB database. Furthermore, feature subsets trained on PTB are validated using the clinical database, resulting in an accuracy of 84.5%. The chosen important features demonstrate that both static and dynamic information have crucial roles in MI detection. Crucially, the proposed method provides clear internal workings in an easy-to-understand visual manner.

A dual-radiomics model for overall survival prediction in early-stage NSCLC patient using pre-treatment CT images

Introduction

Radiation therapy (RT) is one of the primary treatment options for early-stage non-small cell lung cancer (ES-NSCLC). Therefore, accurately predicting the overall survival (OS) rate following radiotherapy is crucial for implementing personalized treatment strategies. This work aims to develop a dual-radiomics (DR) model to (1) predict 3-year OS in ES-NSCLC patients receiving RT using pre-treatment CT images, and (2) provide explanations between feature importanceand model prediction performance.

Methods

The publicly available TCIA Lung1 dataset with 132 ES-NSCLC patients received RT were studied: 89/43 patients in the under/over 3-year OS group. For each patient, two types of radiomic features were examined: 56 handcrafted radiomic features (HRFs) extracted within gross tumor volume, and 512 image deep features (IDFs) extracted using a pre-trained U-Net encoder. They were combined as inputs to an explainable boosting machine (EBM) model for OS prediction. The EBM's mean absolute scores for HRFs and IDFs were used as feature importance explanations. To evaluate identified feature importance, the DR model was compared with EBM using either (1) key or (2) non-key feature type only. Comparison studies with other models, including supporting vector machine (SVM) and random forest (RF), were also included. The performance was evaluated by the area under the receiver operating characteristic curve (AUCROC), accuracy, sensitivity, and specificity with a 100-fold Monte Carlo cross-validation.

Results

The DR model showed highestperformance in predicting 3-year OS (AUCROC=0.81 ± 0.04), and EBM scores suggested that IDFs showed significantly greater importance (normalized mean score=0.0019) than HRFs (score=0.0008). The comparison studies showed that EBM with key feature type (IDFs-only demonstrated comparable AUCROC results (0.81 ± 0.04), while EBM with non-key feature type (HRFs-only) showed limited AUCROC (0.64 ± 0.10). The results suggested that feature importance score identified by EBM is highly correlated with OS prediction performance. Both SVM and RF models were unable to explain key feature type while showing limited overall AUCROC=0.66 ± 0.07 and 0.77 ± 0.06, respectively. Accuracy, sensitivity, and specificity showed a similar trend.

Discussion

In conclusion, a DR model was successfully developed to predict ES-NSCLC OS based on pre-treatment CT images. The results suggested that the feature importance from DR model is highly correlated to the model prediction power.

Explainable Deep Learning-Based Feature Selection and Intrusion Detection Method on the Internet of Things

With the rapid advancement of the Internet of Things, network security has garnered increasing attention from researchers. Applying deep learning (DL) has significantly enhanced the performance of Network Intrusion Detection Systems (NIDSs). However, due to its complexity and "black box" problem, deploying DL-based NIDS models in practical scenarios poses several challenges, including model interpretability and being lightweight. Feature selection (FS) in DL models plays a crucial role in minimizing model parameters and decreasing computational overheads while enhancing NIDS performance. Hence, selecting effective features remains a pivotal concern for NIDSs. In light of this, this paper proposes an interpretable feature selection method for encrypted traffic intrusion detection based on SHAP and causality principles. This approach utilizes the results of model interpretation for feature selection to reduce feature count while ensuring model reliability. We evaluate and validate our proposed method on two public network traffic datasets, CICIDS2017 and NSL-KDD, employing both a CNN and a random forest (RF). Experimental results demonstrate superior performance achieved by our proposed method.

Machine Learning Predicts Patients With New-onset Diabetes at Risk of Pancreatic Cancer

BACKGROUND

New-onset diabetes represent a high-risk cohort to screen for pancreatic cancer.

GOALS

Develop a machine model to predict pancreatic cancer among patients with new-onset diabetes.

STUDY

A retrospective cohort of patients with new-onset diabetes was assembled from multiple health care networks in the United States. An XGBoost machine learning model was designed from a portion of this cohort (the training set) and tested on the remaining part of the cohort (the test set). Shapley values were used to explain the XGBoost's model features. Model performance was compared with 2 contemporary models designed to predict pancreatic cancer among patients with new-onset diabetes.

RESULTS

In the test set, the XGBoost model had an area under the curve of 0.80 (0.76 to 0.85) compared with 0.63 and 0.68 for other models. Using cutoffs based on the Youden index, the sensitivity of the XGBoost model was 75%, the specificity was 70%, the accuracy was 70%, the positive predictive value was 1.2%, and the negative predictive value was >99%. The XGBoost model obtained a positive predictive value of at least 2.5% with a sensitivity of 38%. The XGBoost model was the only model that detected at least 50% of patients with cancer one year after the onset of diabetes. All 3 models had similar features that predicted pancreatic cancer, including older age, weight loss, and the rapid destabilization of glucose homeostasis.

CONCLUSION

Machine learning models isolate a high-risk cohort from those with new-onset diabetes at risk for pancreatic cancer.

LORIS robustly predicts patient outcomes with immune checkpoint blockade therapy using common clinical, pathologic and genomic features

Despite the revolutionary impact of immune checkpoint blockade (ICB) in cancer treatment, accurately predicting patient responses remains challenging. Here, we analyzed a large dataset of 2,881 ICB-treated and 841 non-ICB-treated patients across 18 solid tumor types, encompassing a wide range of clinical, pathologic and genomic features. We developed a clinical score called LORIS (logistic regression-based immunotherapy-response score) using a six-feature logistic regression model. LORIS outperforms previous signatures in predicting ICB response and identifying responsive patients even with low tumor mutational burden or programmed cell death 1 ligand 1 expression. LORIS consistently predicts patient objective response and short-term and long-term survival across most cancer types. Moreover, LORIS showcases a near-monotonic relationship with ICB response probability and patient survival, enabling precise patient stratification. As an accurate, interpretable method using a few readily measurable features, LORIS may help improve clinical decision-making in precision medicine to maximize patient benefit. LORIS is available as an online tool at https://loris.ccr.cancer.gov/ .

Development and Validation of an Interpretable Machine Learning Model for Early Prognosis Prediction in ICU Patients with Malignant Tumors and Hyperkalemia

This study aims to develop and validate a machine learning (ML) predictive model for assessing mortality in patients with malignant tumors and hyperkalemia (MTH). We extracted data on patients with MTH from the Medical Information Mart for Intensive Care-IV, version 2.2 (MIMIC-IV v2.2) database. The dataset was split into a training set (75%) and a validation set (25%). We used the Least Absolute Shrinkage and Selection Operator (LASSO) regression to identify potential predictors, which included clinical laboratory indicators and vital signs. Pearson correlation analysis tested the correlation between predictors. In-hospital death was the prediction target. The Area Under the Curve (AUC) and accuracy of the training and validation sets of 7 ML algorithms were compared, and the optimal 1 was selected to develop the model. The calibration curve was used to evaluate the prediction accuracy of the model further. SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) enhanced model interpretability. 496 patients with MTH in the Intensive Care Unit (ICU) were included. After screening, 17 clinical features were included in the construction of the ML model, and the Pearson correlation coefficient was <0.8, indicating that the correlation between the clinical features was small. eXtreme Gradient Boosting (XGBoost) outperformed other algorithms, achieving perfect scores in the training set (accuracy: 1.000, AUC: 1.000) and high scores in the validation set (accuracy: 0.734, AUC: 0.733). The calibration curves indicated good predictive calibration of the model. SHAP analysis identified the top 8 predictive factors: urine output, mean heart rate, maximum urea nitrogen, minimum oxygen saturation, minimum mean blood pressure, maximum total bilirubin, mean respiratory rate, and minimum pH. In addition, SHAP and LIME performed in-depth individual case analyses. This study demonstrates the effectiveness of ML methods in predicting mortality risk in ICU patients with MTH. It highlights the importance of predictors like urine output and mean heart rate. SHAP and LIME significantly enhanced the model's interpretability.