Interpretable Machine Learning for Choosing Radiation Dose-volume Constraints on Cardio-pulmonary Substructures Associated with Overall Survival in NRG Oncology RTOG 0617.

PURPOSE: Our objective was to use interpretable machine learning for choosing dose-volume constraints on cardiopulmonary substructures (CPSs) associated with overall survival (OS) in radiation therapy for locally advanced non-small cell lung cancer. METHODS AND MATERIALS: A total of 428 patients with non-small cell lung cancer were randomly divided into training/validation/test subsets (n = 230/149/49) in Radiation Therapy Oncology Group 0617. Manual or automated contouring was performed to segment CPSs, including heart, atria, ventricles, aorta, left/right ventricle/atrium (LV+RV+LA+RA), inferior/superior vena cava, pulmonary artery, and pericardium. Peri (pericardium-heart), rest (heart-[LV+RV+LA+RA]), clinical target volume (CTV), and lungs-CTV contours were also obtained. Dose-volume histogram features were extracted, including minimum/mean dose to the hottest x% volume (Dx%[Gy]/MOHx%[Gy]), minimum/mean/maximum dose, percent volume receiving at least xGy (VxGy[%]), and overlapping volume of each CPS with planning target volume (PTV_Voverlap[%]). Clinical parameters were collected from the National Clinical Trials Network/Community oncology research program data archive. Feature selection was performed using a series of multiblock sparse partial least squares regression, stability selection supervised principal component analysis, and Boruta. Explainable boosting machine (EBM) was trained using a conditional survival distribution-based approach for imputing censored data, treating survival analysis as a regression problem. Harrell's C-index was used to evaluate OS discrimination performance of EBM, Cox proportional hazards (CPH), random survival forest, extreme gradient boosting survival embeddings, and CPH deep neural network (DeepSurv) models in the test set. Dose-volume constraints were selected using the binary change point detection algorithm in Shapley additive explanations-based partial dependence functions. RESULTS: Selected features included LA_V60Gy(%), pericardium_D30%(Gy), lungs-CTV_PTV_Voverlap(%), RA_V55Gy(%), and received_cons_chemo. All models ranked LA_V60Gy(%) as the most important feature. EBM achieved the best performance for predicting OS, followed by extreme gradient boosting survival embeddings, random survival forest, DeepSurv, and CPH (C-index = 0.653, 0.646, 0.642, 0.638, and 0.632). EBM global explanations suggested that LA_V60Gy(%) < 25.6, lungs-CTV_PTV_Voverlap(%) < 1.1, pericardium_D30%(Gy) < 18.9, RA_V55Gy(%) < 19.5, and received_cons_chemo = 'Yes' for improved OS. CONCLUSIONS: EBM can be used to discriminate OS while also guiding dose-volume constraint selection for optimal management of cardiac toxicity in lung cancer radiation therapy.

Interpretable machine learning for predicting pathologic complete response in patients treated with chemoradiation therapy for rectal adenocarcinoma.

Purpose: Pathologic complete response (pCR) is a critical factor in determining whether patients with rectal cancer (RC) should have surgery after neoadjuvant chemoradiotherapy (nCRT). Currently, a pathologist's histological analysis of surgical specimens is necessary for a reliable assessment of pCR. Machine learning (ML) algorithms have the potential to be a non-invasive way for identifying appropriate candidates for non-operative therapy. However, these ML models' interpretability remains challenging. We propose using explainable boosting machine (EBM) to predict the pCR of RC patients following nCRT. Methods: A total of 296 features were extracted, including clinical parameters (CPs), dose-volume histogram (DVH) parameters from gross tumor volume (GTV) and organs-at-risk, and radiomics (R) and dosiomics (D) features from GTV. R and D features were subcategorized into shape (S), first-order (L1), second-order (L2), and higher-order (L3) local texture features. Multi-view analysis was employed to determine the best set of input feature categories. Boruta was used to select all-relevant features for each input dataset. ML models were trained on 180 cases from our institution, with 37 cases from RTOG 0822 clinical trial serving as the independent dataset for model validation. The performance of EBM in predicting pCR on the test dataset was evaluated using ROC AUC and compared with that of three state-of-the-art black-box models: extreme gradient boosting (XGB), random forest (RF) and support vector machine (SVM). The predictions of all black-box models were interpreted using Shapley additive explanations. Results: The best input feature categories were CP+DVH+S+R_L1+R_L2 for all models, from which Boruta-selected features enabled the EBM, XGB, RF, and SVM models to attain the AUCs of 0.820, 0.828, 0.828, and 0.774, respectively. Although EBM did not achieve the best performance, it provided the best capability for identifying critical turning points in response scores at distinct feature values, revealing that the bladder with maximum dose >50 Gy, and the tumor with maximum2DDiameterColumn >80 mm, elongation <0.55, leastAxisLength >50 mm and lower variance of CT intensities were associated with unfavorable outcomes. Conclusions: EBM has the potential to enhance the physician's ability to evaluate an ML-based prediction of pCR and has implications for selecting patients for a "watchful waiting" strategy to RC therapy.

Predicting dire outcomes of patients with community acquired pneumonia.

Community-acquired pneumonia (CAP) is an important clinical condition with regard to patient mortality, patient morbidity, and healthcare resource utilization. The assessment of the likely clinical course of a CAP patient can significantly influence decision making about whether to treat the patient as an inpatient or as an outpatient. That decision can in turn influence resource utilization, as well as patient well being. Predicting dire outcomes, such as mortality or severe clinical complications, is a particularly important component in assessing the clinical course of patients. We used a training set of 1601 CAP patient cases to construct 11 statistical and machine-learning models that predict dire outcomes. We evaluated the resulting models on 686 additional CAP-patient cases. The primary goal was not to compare these learning algorithms as a study end point; rather, it was to develop the best model possible to predict dire outcomes. A special version of an artificial neural network (NN) model predicted dire outcomes the best. Using the 686 test cases, we estimated the expected healthcare quality and cost impact of applying the NN model in practice. The particular, quantitative results of this analysis are based on a number of assumptions that we make explicit; they will require further study and validation. Nonetheless, the general implication of the analysis seems robust, namely, that even small improvements in predictive performance for prevalent and costly diseases, such as CAP, are likely to result in significant improvements in the quality and efficiency of healthcare delivery. Therefore, seeking models with the highest possible level of predictive performance is important. Consequently, seeking ever better machine-learning and statistical modeling methods is of great practical significance.

Early statistical detection of anthrax outbreaks by tracking over-the-counter medication sales.

The recent series of anthrax attacks has reinforced the importance of biosurveillance systems for the timely detection of epidemics. This paper describes a statistical framework for monitoring grocery data to detect a large-scale but localized bioterrorism attack. Our system illustrates the potential of data sources that may be more timely than traditional medical and public health data. The system includes several layers, each customized to grocery data and tuned to finding footprints of an epidemic. We also propose an evaluation methodology that is suitable in the absence of data on large-scale bioterrorist attacks and disease outbreaks.