Area Under The Precision-recall Curve Research Articles

Preoperative prediction model for risk of readmission after total joint replacement surgery: a random forest approach leveraging NLP and unfairness mitigation for improved patient care and cost-effectiveness

BackgroundThe Center for Medicare and Medicaid Services (CMS) imposes payment penalties for readmissions following total joint replacement surgeries. This study focuses on total hip, knee, and shoulder arthroplasty procedures as they account for most joint replacement surgeries. Apart from being a burden to healthcare systems, readmissions are also troublesome for patients. There are several studies which only utilized structured data from Electronic Health Records (EHR) without considering any gender and payor bias adjustments.MethodsFor this study, dataset of 38,581 total knee, hip, and shoulder replacement surgeries performed from 2015 to 2021 at Novant Health was gathered. This data was used to train a random forest machine learning model to predict the combined endpoint of emergency department (ED) visit or unplanned readmissions within 30 days of discharge or discharge to Skilled Nursing Facility (SNF) following the surgery. 98 features of laboratory results, diagnoses, vitals, medications, and utilization history were extracted. A natural language processing (NLP) model finetuned from Clinical BERT was used to generate an NLP risk score feature for each patient based on their clinical notes. To address societal biases, a feature bias analysis was performed in conjunction with propensity score matching. A threshold optimization algorithm from the Fairlearn toolkit was used to mitigate gender and payor biases to promote fairness in predictions.ResultsThe model achieved an Area Under the Receiver Operating characteristic Curve (AUROC) of 0.738 (95% confidence interval, 0.724 to 0.754) and an Area Under the Precision-Recall Curve (AUPRC) of 0.406 (95% confidence interval, 0.384 to 0.433). Considering an outcome prevalence of 16%, these metrics indicate the model’s ability to accurately discriminate between readmission and non-readmission cases within the context of total arthroplasty surgeries while adjusting patient scores in the model to mitigate bias based on patient gender and payor.ConclusionThis work culminated in a model that identifies the most predictive and protective features associated with the combined endpoint. This model serves as a tool to empower healthcare providers to proactively intervene based on these influential factors without introducing bias towards protected patient classes, effectively mitigating the risk of negative outcomes and ultimately improving quality of care regardless of socioeconomic factors.

CoMadOut—a robust outlier detection algorithm based on CoMAD

AbstractUnsupervised learning methods are well established in the area of anomaly detection and achieve state of the art performances on outlier datasets. Outliers play a significant role, since they bear the potential to distort the predictions of a machine learning algorithm on a given dataset. Especially among PCA-based methods, outliers have an additional destructive potential regarding the result: they may not only distort the orientation and translation of the principal components, they also make it more complicated to detect outliers. To address this problem, we propose the robust outlier detection algorithm CoMadOut, which satisfies two required properties: (1) being robust towards outliers and (2) detecting them. Our CoMadOut outlier detection variants using comedian PCA define, dependent on its variant, an inlier region with a robust noise margin by measures of in-distribution (variant CMO) and optimized scores by measures of out-of-distribution (variants CMO*), e.g. kurtosis-weighting by CMO+k. These measures allow distribution based outlier scoring for each principal component, and thus, an appropriate alignment of the degree of outlierness between normal and abnormal instances. Experiments comparing CoMadOut with traditional, deep and other comparable robust outlier detection methods showed that the performance of the introduced CoMadOut approach is competitive to well established methods related to average precision (AP), area under the precision recall curve (AUPRC) and area under the receiver operating characteristic (AUROC) curve. In summary our approach can be seen as a robust alternative for outlier detection tasks.

Comparing discriminating abilities of evaluation metrics in link prediction

Abstract Link prediction aims to predict the potential existence of links between two unconnected nodes within a network based on the known topological characteristics. Evaluation metrics are used to assess the effectiveness of algorithms in link prediction. The discriminating ability of these evaluation metrics is vitally important for accurately evaluating link prediction algorithms. In this study, we propose an artificial network model, based on which one can adjust a single parameter to monotonically and continuously turn the prediction accuracy of the specifically designed link prediction algorithm. Building upon this foundation, we show a framework to depict the effectiveness of evaluating metrics by focusing on their discriminating ability. Specifically, a quantitative comparison in the abilities of correctly discerning varying prediction accuracies was conducted encompassing nine evaluation metrics: Precision, Recall, F1-Measure, Matthews Correlation Coefficient (MCC), Balanced Precision (BP), the Area Under the receiver operating characteristic Curve (AUC), the Area Under the Precision-Recall curve (AUPR), Normalized Discounted Cumulative Gain (NDCG), and the Area Under the magnified ROC (AUC-mROC). The results indicate that the discriminating abilities of the three metrics, AUC, AUPR, and NDCG, are significantly higher than those of other metrics.

Prospective validation of a hospital triage predictive model to decrease undertriage: an EAST multicenter study

BackgroundTiered trauma team activation (TTA) allows systems to optimally allocate resources to an injured patient. Target undertriage and overtriage rates of <5% and <35% are difficult for centers to achieve,...

Development and external validation of deep learning clinical prediction models using variable-length time series data.

To compare and externally validate popular deep learning model architectures and data transformation methods for variable-length time series data in 3 clinical tasks (clinical deterioration, severe acute kidney injury [AKI], and suspected infection). This multicenter retrospective study included admissions at 2 medical centers that spanned 2007-2022. Distinct datasets were created for each clinical task, with 1 site used for training and the other for testing. Three feature engineering methods (normalization, standardization, and piece-wise linear encoding with decision trees [PLE-DTs]) and 3 architectures (long short-term memory/gated recurrent unit [LSTM/GRU], temporal convolutional network, and time-distributed wrapper with convolutional neural network [TDW-CNN]) were compared in each clinical task. Model discrimination was evaluated using the area under the precision-recall curve (AUPRC) and the area under the receiver operating characteristic curve (AUROC). The study comprised 373825 admissions for training and 256128 admissions for testing. LSTM/GRU models tied with TDW-CNN models with both obtaining the highest mean AUPRC in 2 tasks, and LSTM/GRU had the highest mean AUROC across all tasks (deterioration: 0.81, AKI: 0.92, infection: 0.87). PLE-DT with LSTM/GRU achieved the highest AUPRC in all tasks. When externally validated in 3 clinical tasks, the LSTM/GRU model architecture with PLE-DT transformed data demonstrated the highest AUPRC in all tasks. Multiple models achieved similar performance when evaluated using AUROC. The LSTM architecture performs as well or better than some newer architectures, and PLE-DT may enhance the AUPRC in variable-length time series data for predicting clinical outcomes during external validation.

Interpretable machine learning in predicting drug-induced liver injury among tuberculosis patients: model development and validation study.

The objective of this research was to create and validate an interpretable prediction model for drug-induced liver injury (DILI) during tuberculosis (TB) treatment. A dataset of TB patients from Ningbo City was used to develop models employing the eXtreme Gradient Boosting (XGBoost), random forest (RF), and the least absolute shrinkage and selection operator (LASSO) logistic algorithms. The model's performance was evaluated through various metrics, including the area under the receiver operating characteristic curve (AUROC) and the area under the precision recall curve (AUPR) alongside the decision curve. The Shapley Additive exPlanations (SHAP) method was used to interpret the variable contributions of the superior model. A total of 7,071TB patients were identified from the regional healthcare dataset. The study cohort consisted of individuals with a median age of 47years, 68.0% of whom were male, and 16.3% developed DILI. We utilized part of the high dimensional propensity score (HDPS) method to identify relevant variables and obtained a total of 424 variables. From these, 37 variables were selected for inclusion in a logistic model using LASSO. The dataset was then split into training and validation sets according to a 7:3 ratio. In the validation dataset, the XGBoost model displayed improved overall performance, with an AUROC of 0.89, an AUPR of 0.75, an F1 score of 0.57, and a Brier score of 0.07. Both SHAP analysis and XGBoost model highlighted the contribution of baseline liver-related ailments such as DILI, drug-induced hepatitis (DIH), and fatty liver disease (FLD). Age, alanine transaminase (ALT), and total bilirubin (Tbil) were also linked to DILI status. XGBoost demonstrates improved predictive performance compared to RF and LASSO logistic in this study. Moreover, the introduction of the SHAP method enhances the clinical understanding and potential application of the model. For further research, external validation and more detailed feature integration are necessary.

Predicting haemoglobin deferral using machine learning models: Can we use the same prediction model across countries?

Personalized donation strategies based on haemoglobin (Hb) prediction models may reduce Hb deferrals and hence costs of donation, meanwhile improving commitment of donors. We previously found that prediction models perform better in validation data with a high Hb deferral rate. We therefore investigate how Hb deferral prediction models perform when exchanged with other blood establishments. Donation data from the past 5 years from random samples of 10,000 donors from Australia, Belgium, Finland, the Netherlands and South Africa were used to fit random forest models for Hb deferral prediction. Trained models were exchanged between blood establishments. Model performance was evaluated using the area under the precision-recall curve (AUPR). Variable importance was assessed using SHapley Additive exPlanations (SHAP) values. Across the validation datasets and exchanged models, the AUPR ranged from 0.05 to 0.43. Exchanged models performed similarly within validation datasets, irrespective of the origin of the training data. Apart from subtle differences, the importance of most predictor variables was similar in all trained models. Our results suggest that Hb deferral prediction models trained in different blood establishments perform similarly within different validation datasets, regardless of the deferral rate of their training data. Models learn similar associations in different blood establishments.

Machine Learning Algorithms to Predict Colistin-Induced Nephrotoxicity from Electronic Health Records in Patients with Multidrug-Resistant Gram-Negative Infection.

Colistin-induced nephrotoxicity prolongs hospitalization and increases mortality. The study aimed to construct machine learning models to predict colistin-induced nephrotoxicity in patients with multidrug-resistant gram-negative infection. Patients receiving colistin from three hospitals in the Clinical Research Database were included. Data were divided into a derivation cohort (2011∼2017) and a temporal validation cohort (2018∼2020). Fifteen machine learning models were established by categorical boosting, light gradient boosting machine, and random forest. Classifier performances were compared by the sensitivity, F1 score, Matthews correlation coefficient (MCC), area under the receiver operating characteristic (AUROC) curve, and area under the precision-recall curve (AUPRC). SHapley Additive exPlanations plots were drawn to understand feature importance and interactions. The study included 1392 patients, with 360 (36.4%) and 165 (40.9%) experiencing nephrotoxicity in the derivation and temporal validation cohorts, respectively. The categorical boosting with oversampling achieved the highest performance with a sensitivity of 0.860, an F1 score of 0.740, an MCC of 0.533, an AUROC curve of 0.823, and an AUPRC of 0.737. The feature importance demonstrated that the days of colistin use, cumulative dose, daily dose, latest C-reactive protein, and baseline hemoglobin were the most important risk factors, especially for vulnerable patients. A cutoff colistin dose of 4.0 mg/kg body weight/day was identified for patients at higher risk of nephrotoxicity. Machine learning techniques can be an early identification tool to predict colistin-induced nephrotoxicity. The observed interactions suggest a modification in dose adjustment guidelines. Future geographic and prospective validation studies are warranted to strengthen the real-world applicability.

Machine learning models predict the emergence of depression in Argentinean college students during periods of COVID-19 quarantine.

The COVID-19 pandemic has exacerbated mental health challenges, particularly depression among college students. Detecting at-risk students early is crucial but remains challenging, particularly in developing countries. Utilizing data-driven predictive models presents a viable solution to address this pressing need. 1) To develop and compare machine learning (ML) models for predicting depression in Argentinean students during the pandemic. 2) To assess the performance of classification and regression models using appropriate metrics. 3) To identify key features driving depression prediction. A longitudinal dataset (N = 1492 college students) captured T1 and T2 measurements during the Argentinean COVID-19 quarantine. ML models, including linear logistic regression classifiers/ridge regression (LogReg/RR), random forest classifiers/regressors, and support vector machines/regressors (SVM/SVR), are employed. Assessed features encompass depression and anxiety scores (at T1), mental disorder/suicidal behavior history, quarantine sub-period information, sex, and age. For classification, models' performance on test data is evaluated using Area Under the Precision-Recall Curve (AUPRC), Area Under the Receiver Operating Characteristic curve, Balanced Accuracy, F1 score, and Brier loss. For regression, R-squared (R2), Mean Absolute Error, and Mean Squared Error are assessed. Univariate analyses are conducted to assess the predictive strength of each individual feature with respect to the target variable. The performance of multi- vs univariate models is compared using the mean AUPRC score for classifiers and the R2 score for regressors. The highest performance is achieved by SVM and LogReg (e.g., AUPRC: 0.76, 95% CI: 0.69, 0.81) and SVR and RR models (e.g., R2 for SVR and RR: 0.56, 95% CI: 0.45, 0.64 and 0.45, 0.63, respectively). Univariate models, particularly LogReg and SVM using depression (AUPRC: 0.72, 95% CI: 0.64, 0.79) or anxiety scores (AUPRC: 0.71, 95% CI: 0.64, 0.78) and RR using depression scores (R2: 0.48, 95% CI: 0.39, 0.57) exhibit performance levels close to those of the multivariate models, which include all features. These findings highlight the relevance of pre-existing depression and anxiety conditions in predicting depression during quarantine, underscoring their comorbidity. ML models, particularly SVM/SVR and LogReg/RR, demonstrate potential in the timely detection of at-risk students. However, further studies are needed before clinical implementation.

A novel electronic health record-based, machine-learning model to predict severe hypoglycemia leading to hospitalizations in older adults with diabetes: A territory-wide cohort and modeling study.

Older adults with diabetes are at high risk of severe hypoglycemia (SH). Many machine-learning (ML) models predict short-term hypoglycemia are not specific for older adults and show poor precision-recall. We aimed to develop a multidimensional, electronic health record (EHR)-based ML model to predict one-year risk of SH requiring hospitalization in older adults with diabetes. We adopted a case-control design for a retrospective territory-wide cohort of 1,456,618 records from 364,863 unique older adults (age ≥65 years) with diabetes and at least 1 Hong Kong Hospital Authority attendance from 2013 to 2018. We used 258 predictors including demographics, admissions, diagnoses, medications, and routine laboratory tests in a one-year period to predict SH events requiring hospitalization in the following 12 months. The cohort was randomly split into training, testing, and internal validation sets in a 7:2:1 ratio. Six ML algorithms were evaluated including logistic-regression, random forest, gradient boost machine, deep neural network (DNN), XGBoost, and Rulefit. We tested our model in a temporal validation cohort in the Hong Kong Diabetes Register with predictors defined in 2018 and outcome events defined in 2019. Predictive performance was assessed using area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC) statistics, and positive predictive value (PPV). We identified 11,128 SH events requiring hospitalization during the observation periods. The XGBoost model yielded the best performance (AUROC = 0.978 [95% CI 0.972 to 0.984]; AUPRC = 0.670 [95% CI 0.652 to 0.688]; PPV = 0.721 [95% CI 0.703 to 0.739]). This was superior to an 11-variable conventional logistic-regression model comprised of age, sex, history of SH, hypertension, blood glucose, kidney function measurements, and use of oral glucose-lowering drugs (GLDs) (AUROC = 0.906; AUPRC = 0.085; PPV = 0.468). Top impactful predictors included non-use of lipid-regulating drugs, in-patient admission, urgent emergency triage, insulin use, and history of SH. External validation in the HKDR cohort yielded AUROC of 0.856 [95% CI 0.838 to 0.873]. Main limitations of this study included limited transportability of the model and lack of geographically independent validation. Our novel-ML model demonstrated good discrimination and high precision in predicting one-year risk of SH requiring hospitalization. This may be integrated into EHR decision support systems for preemptive intervention in older adults at highest risk.

Deep learning to predict rapid progression of Alzheimer's disease from pooled clinical trials: A retrospective study.

The rate of progression of Alzheimer's disease (AD) differs dramatically between patients. Identifying the most is critical because when their numbers differ between treated and control groups, it distorts the outcome, making it impossible to tell whether the treatment was beneficial. Much recent effort, then, has gone into identifying RPs. We pooled de-identified placebo-arm data of three randomized controlled trials (RCTs), EXPEDITION, EXPEDITION 2, and EXPEDITION 3, provided by Eli Lilly and Company. After processing, the data included 1603 mild-to-moderate AD patients with 80 weeks of longitudinal observations on neurocognitive health, brain volumes, and amyloid-beta (Aβ) levels. RPs were defined by changes in four neurocognitive/functional health measures. We built deep learning models using recurrent neural networks with attention mechanisms to predict RPs by week 80 based on varying observation periods from baseline (e.g., 12, 28 weeks). Feature importance scores for RP prediction were computed and temporal feature trajectories were compared between RPs and non-RPs. Our evaluation and analysis focused on models trained with 28 weeks of observation. The models achieved robust internal validation area under the receiver operating characteristic (AUROCs) ranging from 0.80 (95% CI 0.79-0.82) to 0.82 (0.81-0.83), and the area under the precision-recall curve (AUPRCs) from 0.34 (0.32-0.36) to 0.46 (0.44-0.49). External validation AUROCs ranged from 0.75 (0.70-0.81) to 0.83 (0.82-0.84) and AUPRCs from 0.27 (0.25-0.29) to 0.45 (0.43-0.48). Aβ plasma levels, regional brain volumetry, and neurocognitive health emerged as important factors for the model prediction. In addition, the trajectories were stratified between predicted RPs and non-RPs based on factors such as ventricular volumes and neurocognitive domains. Our findings will greatly aid clinical trialists in designing tests for new medications, representing a key step toward identifying effective new AD therapies.

Explainable Machine Learning Model to Preoperatively Predict Postoperative Complications in Inpatients With Cancer Undergoing Major Operations.

Preoperative prediction of postoperative complications (PCs) in inpatients with cancer is challenging. We developed an explainable machine learning (ML) model to predict PCs in a heterogenous population of inpatients with cancer undergoing same-hospitalization major operations. Consecutive inpatients who underwent same-hospitalization operations from December 2017 to June 2021 at a single institution were retrospectively reviewed. The ML model was developed and tested using electronic health record (EHR) data to predict 30-day PCs for patients with Clavien-Dindo grade 3 or higher (CD 3+) per the CD classification system. Model performance was assessed using area under the receiver operating characteristic curve (AUROC), area under the precision recall curve (AUPRC), and calibration plots. Model explanation was performed using the Shapley additive explanations (SHAP) method at cohort and individual operation levels. A total of 988 operations in 827 inpatients were included. The ML model was trained using 788 operations and tested using a holdout set of 200 operations. The CD 3+ complication rates were 28.6% and 27.5% in the training and holdout test sets, respectively. Training and holdout test sets' model performance in predicting CD 3+ complications yielded an AUROC of 0.77 and 0.73 and an AUPRC of 0.56 and 0.52, respectively. Calibration plots demonstrated good reliability. The SHAP method identified features and the contributions of the features to the risk of PCs. We trained and tested an explainable ML model to predict the risk of developing PCs in patients with cancer. Using patient-specific EHR data, the ML model accurately discriminated the risk of developing CD 3+ complications and displayed top features at the individual operation and cohort level.

VEpiNet: A multimodal interictal epileptiform discharge detection method based on video and electroencephalogram data

To enhance deep learning-based automated interictal epileptiform discharge (IED) detection, this study proposes a multimodal method, vEpiNet, that leverages video and electroencephalogram (EEG) data. Datasets comprise 24 931 IED (from 484 patients) and 166 094 non-IED 4-second video-EEG segments. The video data is processed by the proposed patient detection method, with frame difference and Simple Keypoints (SKPS) capturing patients’ movements. EEG data is processed with EfficientNetV2. The video and EEG features are fused via a multilayer perceptron. We developed a comparative model, termed nEpiNet, to test the effectiveness of the video feature in vEpiNet. The 10-fold cross-validation was used for testing. The 10-fold cross-validation showed high areas under the receiver operating characteristic curve (AUROC) in both models, with a slightly superior AUROC (0.9902) in vEpiNet compared to nEpiNet (0.9878). Moreover, to test the model performance in real-world scenarios, we set a prospective test dataset, containing 215 h of raw video-EEG data from 50 patients. The result shows that the vEpiNet achieves an area under the precision–recall curve (AUPRC) of 0.8623, surpassing nEpiNet’s 0.8316. Incorporating video data raises precision from 70% (95% CI, 69.8%–70.2%) to 76.6% (95% CI, 74.9%–78.2%) at 80% sensitivity and reduces false positives by nearly a third, with vEpiNet processing one-hour video-EEG data in 5.7 min on average. Our findings indicate that video data can significantly improve the performance and precision of IED detection, especially in prospective real clinic testing. It suggests that vEpiNet is a clinically viable and effective tool for IED analysis in real-world applications.

Clinical Features Predicting COVID-19 Severity Risk at the Time of Hospitalization.

The global spread of COVID-19 has led to significant mortality and morbidity worldwide. Early identification of COVID-19 patients who are at high risk of developing severe disease can help in improved patient management, care, and treatment, as well as in the effective allocation of hospital resources. The severity prediction at the time of hospitalization can be extremely helpful in deciding the treatment of COVID-19 patients. To this end, this study presents an interpretable artificial intelligence (AI) model, named COVID-19 severity predictor (CoSP) that predicts COVID-19 severity using the clinical features at the time of hospital admission. We utilized a dataset comprising 64 demographic and laboratory features of 7,416 confirmed COVID-19 patients that were collected at the time of hospital admission. The proposed hierarchical CoSP model performs four-class COVID severity risk prediction into asymptomatic, mild, moderate, and severe categories. CoSP yielded better performance with good interpretability, as observed via Shapley analysis on COVID severity prediction compared to the other popular ML methods, with an area under the received operating characteristic curve (AUC-ROC) of 0.95, an area under the precision-recall curve (AUPRC) of 0.91, and a weighted F1-score of 0.83. Out of 64 initial features, 19 features were inferred as predictive of the severity of COVID-19 disease by the CoSP model. Therefore, an AI model predicting COVID-19 severity may be helpful for early intervention, optimizing resource allocation, and guiding personalized treatments, potentially enabling healthcare professionals to save lives and allocate resources effectively in the fight against the pandemic.

Clinical performance of automated machine learning: A systematic review

Introduction: Automated machine learning (autoML) removes technical and technological barriers to building artificial intelligence models. We aimed to summarise the clinical applications of autoML, assess the capabilities of utilised platforms, evaluate the quality of the evidence trialling autoML, and gauge the performance of autoML platforms relative to conventionally developed models, as well as each other. Method: This review adhered to a prospectively registered protocol (PROSPERO identifier CRD42022344427). The Cochrane Library, Embase, MEDLINE and Scopus were searched from inception to 11 July 2022. Two researchers screened abstracts and full texts, extracted data and conducted quality assessment. Disagreement was resolved through discussion and as if required, arbitration by a third researcher. Results: There were 26 distinct autoML platforms featured in 82 studies. Brain and lung disease were the most common fields of study of 22 specialties. AutoML exhibited variable performance: area under the receiver operator characteristic curve (AUCROC) 0.35–1.00, F1-score 0.16–0.99, area under the precision-recall curve (AUPRC) 0.51–1.00. AutoML exhibited the highest AUCROC in 75.6% trials; the highest F1-score in 42.3% trials; and the highest AUPRC in 83.3% trials. In autoML platform comparisons, AutoPrognosis and Amazon Rekognition performed strongest with unstructured and structured data, respectively. Quality of reporting was poor, with a median DECIDE-AI score of 14 of 27. Conclusion: A myriad of autoML platforms have been applied in a variety of clinical contexts. The performance of autoML compares well to bespoke computational and clinical benchmarks. Further work is required to improve the quality of validation studies. AutoML may facilitate a transition to data-centric development, and integration with large language models may enable AI to build itself to fulfil user-defined goals.

Predicting Successful Weaning from Mechanical Ventilation by Reduction in Positive End-expiratory Pressure Level Using Machine Learning.

Weaning patients from mechanical ventilation (MV) is a critical and resource intensive process in the Intensive Care Unit (ICU) that impacts patient outcomes and healthcare expenses. Weaning methods vary widely among providers. Prolonged MV is associated with adverse events and higher healthcare expenses. Predicting weaning readiness is a non-trivial process in which the positive end-expiratory pressure (PEEP), a crucial component of MV, has potential to be indicative but has not yet been used as the target. We aimed to predict successful weaning from mechanical ventilation by targeting changes in the PEEP-level using a supervised machine learning model. This retrospective study included 12,153 mechanically ventilated patients from Medical Information Mart for Intensive Care (MIMIC-IV) and eICU collaborative research database (eICU-CRD). Two machine learning models (Extreme Gradient Boosting and Logistic Regression) were developed using a continuous PEEP reduction as target. The data is splitted into 80% as training set and 20% as test set. The model's predictive performance was reported using 95% confidence interval (CI), based on evaluation metrics such as area under the receiver operating characteristic (AUROC), area under the precision-recall curve (AUPRC), F1-Score, Recall, positive predictive value (PPV), and negative predictive value (NPV). The model's descriptive performance was reported as the variable ranking using SHAP (SHapley Additive exPlanations) algorithm. The best model achieved an AUROC of 0.84 (95% CI 0.83-0.85) and an AUPRC of 0.69 (95% CI 0.67-0.70) in predicting successful weaning based on the PEEP reduction. The model demonstrated a Recall of 0.85 (95% CI 0.84-0.86), F1-score of 0.86 (95% CI 0.85-0.87), PPV of 0.87 (95% CI 0.86-0.88), and NPV of 0.64 (95% CI 0.63-0.66). Most of the variables that SHAP algorithm ranked to be important correspond with clinical intuition, such as duration of MV, oxygen saturation (SaO2), PEEP, and Glasgow Coma Score (GCS) components. This study demonstrates the potential application of machine learning in predicting successful weaning from MV based on continuous PEEP reduction. The model's high PPV and moderate NPV suggest that it could be a useful tool to assist clinicians in making decisions regarding ventilator management.

Clinical performance of automated machine learning: A systematic review

Introduction: Automated machine learning (autoML) removes technical and technological barriers to building artificial intelligence models. We aimed to summarise the clinical applications of autoML, assess the capabilities of utilised platforms, evaluate the quality of the evidence trialling autoML, and gauge the performance of autoML platforms relative to conventionally developed models, as well as each other. Method: This review adhered to a prospectively registered protocol (PROSPERO identifier CRD42022344427). The Cochrane Library, Embase, MEDLINE and Scopus were searched from inception to 11 July 2022. Two researchers screened abstracts and full texts, extracted data and conducted quality assessment. Disagreement was resolved through discussion and as if required, arbitration by a third researcher. Results: There were 26 distinct autoML platforms featured in 82 studies. Brain and lung disease were the most common fields of study of 22 specialties. AutoML exhibited variable performance: area under the receiver operator characteristic curve (AUCROC) 0.35–1.00, F1-score 0.16–0.99, area under the precision-recall curve (AUPRC) 0.51–1.00. AutoML exhibited the highest AUCROC in 75.6% trials; the highest F1-score in 42.3% trials; and the highest AUPRC in 83.3% trials. In autoML platform comparisons, AutoPrognosis and Amazon Rekognition performed strongest with unstructured and structured data, respectively. Quality of reporting was poor, with a median DECIDE-AI score of 14 of 27. Conclusion: A myriad of autoML platforms have been applied in a variety of clinical contexts. The performance of autoML compares well to bespoke computational and clinical benchmarks. Further work is required to improve the quality of validation studies. AutoML may facilitate a transition to data-centric development, and integration with large language models may enable AI to build itself to fulfil user-defined goals.

A comparative evaluation of machine learning ensemble approaches for disease prediction using multiple datasets

PurposeMachine learning models are used to develop and improve various disease prediction systems. Ensemble learning is a machine learning technique that combines many classifiers to increase performance by making more accurate predictions than a single classifier. Although several researchers have employed ensemble techniques for disease prediction, a comprehensive comparative study of these techniques still needs to be provided.MethodsUsing 16 disease datasets from Kaggle and the UCI Machine Learning Repository, this study compares the performance of 15 variants of ensemble techniques for disease prediction. The comparison was performed using six performance measures: accuracy, precision, recall, F1 score, AUC (Area Under the receiver operating characteristics Curve) and AUPRC (Area Under the Precision-Recall Curve).ResultsStacking variant of Multi-level stacking showed superior disease prediction performance compared with other bagging and boosting variants, followed by another stacking variant (Classical stacking). Overall, stacking outperformed bagging and boosting for disease prediction. Logit Boost showed the worst performance.ConclusionThe findings of this study can help researchers select an appropriate ensemble approach for future studies focusing on accurate disease prediction.

Feature selection strategies: a comparative analysis of SHAP-value and importance-based methods

In the context of high-dimensional credit card fraud data, researchers and practitioners commonly utilize feature selection techniques to enhance the performance of fraud detection models. This study presents a comparison in model performance using the most important features selected by SHAP (SHapley Additive exPlanations) values and the model’s built-in feature importance list. Both methods rank features and choose the most significant ones for model assessment. To evaluate the effectiveness of these feature selection techniques, classification models are built using five classifiers: XGBoost, Decision Tree, CatBoost, Extremely Randomized Trees, and Random Forest. The Area under the Precision-Recall Curve (AUPRC) serves as the evaluation metric. All experiments are executed on the Kaggle Credit Card Fraud Detection Dataset. The experimental outcomes and statistical tests indicate that feature selection methods based on importance values outperform those based on SHAP values across classifiers and various feature subset sizes. For models trained on larger datasets, it is recommended to use the model’s built-in feature importance list as the primary feature selection method over SHAP. This suggestion is based on the rationale that computing SHAP feature importance is a distinct activity, while models naturally provide built-in feature importance as part of the training process, requiring no additional effort. Consequently, opting for the model’s built-in feature importance list can offer a more efficient and practical approach for larger datasets and more intricate models.

Abstract 890: Multimodal modeling of digitized histopathology slides improves risk stratification in hormone receptor-positive breast cancer patients

Abstract In early-stage hormone receptor-positive breast cancer, genomic risk scores identify patients who stand to benefit from up-front chemotherapy but introduce financial and logistical hurdles to care. We assembled a cohort of 5,244 patients with 11,671 corresponding whole-side images of breast tumors stained with hematoxylin and eosin. We developed a multimodal machine learning model to infer risk of distal metastatic recurrence from routine clinical data. Specifically, the model interprets text from the pathologist’s report using a large language model and uses self-supervised vision transformers to interpret the corresponding whole-slide image. Tensor fusion joins the modalities to infer Genomic Health’s Oncotype DX recurrence score. Inferred recurrence score from the multimodal model correlated with measured score with a concordance correlation coefficient of 0.64 (95% C.I. 0.59 - 0.69) in the withheld test set, compared to 0.55 (95% C.I. 0.49 - 0.61) and 0.56 (95% C.I. 0.52 - 0.60) for the linguistic and visual unimodal models, respectively. The multimodal model attains an area under the precision-recall curve (AUPRC) of 0.69 (AUROC=0.88) for identifying high-risk disease in the full-information setting (when images and pathology reports with quantitative hormone receptor status and grade are available) in a withheld test set, compared to AUPRC of 0.61 and 0.66 for the linguistic and visual models, respectively. By comparison, in the same full-information setting, the clinical nomogram introduced by Orucevic et al. in 2019 achieves an AUPRC of 0.48. We suggest the operating point at which precision is 94.4% and recall is 33.3%. Digitized whole-slide images of routine breast biopsies and their associated synoptic pathology reports contain much of the information necessary to stratify patients by risk of distal metastatic recurrence, when modeled appropriately. Our model could enable hospitals to rapidly triage the need for genomic risk testing, possibly precluding one third of orders without loss of accuracy. This helps allocate scarce resources for genomic tests and valuable weeks prior to beginning therapy while maintaining the standard of precision oncology. Citation Format: Kevin M. Boehm, Antonio Marra, Jorge S. Reis-Filho, Sarat Chandarlapaty, Fresia Pareja, Sohrab P. Shah. Multimodal modeling of digitized histopathology slides improves risk stratification in hormone receptor-positive breast cancer patients [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 890.