eScholarship@UMassChan
eScholarship@UMassChan is a digital archive for UMass Chan Medical School's research and scholarship, including journal articles, theses, datasets and more. We welcome submissions from our faculty, staff, and students. eScholarship@UMassChan is a service of the Lamar Soutter Library, Worcester, MA, USA. See also our open access journal publishing services.
Questions? See the Help menu or contact escholarship@umassmed.edu.
Communities in eScholarship@UMassChan
Select a community to browse its collections.
Recent Publications
Publication Integrating Symbolic Regression for Generalizable and Interpretable Machine Learning in Cardiovascular Risk Prediction(2025-11-18)Introduction: Cardiovascular disease (CVD) risk prediction is crucial for timely, targeted risk factor modification. Machine learning (ML) is often applied to CVD risk prediction to improve diagnosis and treatment but with modest performance [1] and limited interpretability - often requiring large datasets with a risk of overfitting. Enhancing the generalizability and interpretability of CVD risk prediction ML models is critical for their broad adoption and clinical deployment. We explored the integration of symbolic regression (SR) with random forests (RF) to address this challenge. Methods: We analyzed 518,389 patients in the UMass Memorial Clarity data lake research database. These patients had at least one clinical encounter between Oct 1, 2017 and Nov 1, 2018, without any indication of death or CVD during that time based on ICD-10 codes. [2] Demographic features (e.g., age, race/ethnicity) and known risk factors (e.g., diagnoses, lab results) were extracted from the electronic health records. Pre-processing included examination of missing/spurious data, imputation using MissForest, one-hot encoding, and aggregation of laboratory and diagnostic features by condition and measured quantity. We developed and validated a gpu-accelerated SR-enhanced RF (SReRF), motivated by prior work hybridizing SR with decision trees, [3] to predict 5-year CVD risk. Performance metrics included precision, recall, F1, area under the ROC curve (AUROC), and area under the precision-recall curve (AUPRC). We did a 75-25 split for training/testing data and tuned an RF using cross validation and grid search. Results: The patient cohort had a mean (SD) age of 48.6 (18.7), with 5.8% CVD events observed in the 5-year window. SReRF yielded superior performance compared with classical RF. Specifically, classical RF had an AUROC (SD) of 0.79 (0.0027), AUPRC 0.18 (0.0027), and F1 0.27 (0.0032), compared with SReRF’s AUROC of 0.82 (0.0023), AUPRC 0.20 (0.0034), and F1 0.29 (0.0034). Here, SDs were computed using 500 bootstrapped samples of the test set. Visual examination of PR curves indicated that the SReRF’s performance improvement was substantial for thresholds associated with lower recall values, but modest for higher recall thresholds. Also, whereas the classical RF overfit the training set, the SReRF had nearly identical performance on both training and test sets. The main predictors in SReRF (higher in the decision trees) were age, HDL and LDL cholesterol, hypertension, atrial fibrillation, chronic obstructive pulmonary disease, obesity, albumin, creatinine, and self-reported Hispanic/Latinx ethnicity. Discussion and Conclusion: This study demonstrates promising results of integrating symbolic regression with random forest for CVD risk prediction. We found that SReRF outperformed classic RF. The top SReRF predictors align with known risk factors for CVD, adding face validity and reinforcing its potential clinical utility. Further, the symbolic expressions that determined splits produced compact and meaningful relationships between predictors, improving model interpretability. Expressions revealed the simultaneous presence of related comorbidities (e.g., hyperlipidemia and hypertension) and non-linear relationships between numerical predictors (e.g. age and LDL). Of note, ethnicity (Hispanic/Latinx) among identified main predictors may reflect underlying disparities in social determinants of health or biological risk factors. Despite recognized advantages of SR, limitations exist, including long training time, difficulty handling missing data, and limited scalability with high-dimension features. Future models will address these issues, include enhanced pre-processing, computational optimization, additional clinical features (e.g. medications), and be evaluated against other contemporary ML techniques. In conclusion, SReRF represents a promising, interpretable, and generalizable approach for CVD risk prediction, with strong potential for implementation in clinical settings, especially those with a high demand for transparency in predictive modeling. References 1. Soares C, Kwok M, Boucher K, et al. Performance of Cardiovascular Risk Prediction Models Among People Living With HIV: A Systematic Review and Meta-analysis. JAMA Cardiol. 2023;8(2):139–149. 2. https://vsac.nlm.nih.gov/valueset/2.16.840.1.113762.1.4.1078.90/expansion/Latest 3. Fong, Kei Sen, and Mehul Motani. 2024. “Symbolic Regression Enhanced Decision Trees for Classification Tasks.” Proceedings of the AAAI Conference on Artificial Intelligence 38 (11): 12033–42.Publication Patients Who Screen Positive for Suicide Risk Incidental to Their Chief Complaint in the Emergency Department: Characteristics and Post-Visit Suicide Outcomes(2025-11-07)Objective: Universal screening improves suicide risk detection in individuals presenting to the emergency department (ED) who are not presenting with a psychiatric chief complaint, what we refer to as incidental risk. We sought to better understand characteristics of individuals who present with incidental risk and to evaluate their suicide-related outcomes after the ED visit. Methods: Two samples (cross-sectional, longitudinal) from the Emergency Department Safety Assessment and Follow-up Evaluation (ED-SAFE) study were used. Combined, the samples allowed for comparison of baseline characteristics and suicide-related outcomes for participants with incidental risk compared to those with negligible risk (any kind of chief complaint and negative suicide risk) and clinically congruent risk (psychiatric chief complaint and positive suicide risk). Univariable and multivariable logistic regression analyses were completed. Results: Universal screening differentially improved the identification of suicide risk among non-white individuals, potentially reducing racial disparities in risk detection. Participants presenting with incidental risk were generally more similar to those with congruent risk than they were to those with negligible risk across demographics and clinical characteristics. Those with incidental suicide risk exhibited similar post-visit suicide-related outcomes compared to those with congruent risk, yet they were far less likely to receive clinical assessments and interventions during the ED visit. Conclusions: The results of this study highlight an opportunity to broaden evidence-based suicide prevention practices in the ED where logistically possible. EDs may need to consider redesigning their clinical approach to address suicide risk among those who present with medical complaints but screen positive for suicide risk.Publication Critical review of partial volume correction methods in PET and SPECT imaging: benefits, pitfalls, challenges, and future outlook(2025-11-05)Purpose: Partial volume effects (PVE) remain a major challenge in quantitative single-photon emission computed tomography (SPECT) and positron emission tomography (PET) imaging, often compromising both accuracy and reproducibility. While numerous Partial Volume Correction (PVC) methods have been proposed, their clinical translation is still limited. This review provides a clinically oriented evaluation of PVC methods with a particular focus on state-of-the-art applications in neurology, cardiovascular imaging, oncology, and radiopharmaceutical therapy dosimetry, highlighting where these techniques offer the greatest added value. In addition, we outline which PVC techniques have the potential to be used in clinical practice and which remain primarily suited for research purposes, along with their suitability in each of the above-mentioned clinical domains. Finally, this review addresses the central question of whether PVC is essential in clinical practice or whether its impact is context dependent. Methods: This review categorizes PVC approaches into three partially overlapping classes: reconstruction-based, post-reconstruction-based, and AI-driven or hybrid methods. Each class is further divided into anatomical and non-anatomical subcategories. We systematically compare their clinical applicability across key dimensions: quantitative accuracy, lesion detectability, robustness to noise and artifacts, anatomical dependence, generalizability across scanners and tracers, and clinical readiness. Results: PVC techniques often improve quantitative accuracy in small structures and in regions affected by spill-over from adjacent high-uptake tissues. However, these benefits can come at the cost of increased noise or edge artifacts, which may limit their robustness for routine clinical use. Post-reconstruction methods are sensitive to segmentation errors, while AI-driven models, despite their promise, require further validation using clinical benchmarks, comparison to ground truth, and testing on diverse datasets. Issues, such as generalizability and interpretability remain significant barriers. Conclusion: This review emphasizes the importance of application-tailored PVC protocols for reliable quantitative imaging in neurology, cardiology, oncology, and radiopharmaceutical therapy dosimetry. Not all PVC methods are beneficial; some may even impair interpretation in certain contexts. We provide a practical overview of which PVC approaches are most beneficial for each clinical scenario, aiming to guide both researchers and clinicians in selecting appropriate techniques for future studies and routine practice, and also outline key areas requiring further development for broader integration into research and clinical workflows.Publication Adverse cardiometabolic impacts of sleep fragmentation and estradiol suppression: An experimental model of menopause(2025-11-05)Context: Risk of cardiometabolic disease increases in women transitioning to postmenopause, during which estradiol declines universally. Most of these women experience fragmentation of sleep due to nocturnal hot flashes, without a reduction in total sleep time. Objective: We examined the independent impact of estradiol suppression, sleep, and their combination on cardiometabolic outcomes categorized as satiety and hunger, lipid profile, cardiac vital signs, and glucoregulation. Design: Participants completed 5-night inpatient studies under eucaloric conditions, once during mid-follicular phase/estrogenized and again under estrogen suppressed conditions, using the same experimental protocol both times. For all participants, sleep was unfragmented the first two nights and then experimentally fragmented without reducing total sleep time the next three nights. Setting: Inpatient Intensive Physiological Monitoring research facility. Participants: 38 healthy premenopausal women. Intervention(s): Clinical experimental induced menopause model including gonadotropin-releasing hormone agonist-induced hypoestrogenism and sleep fragmentation. Main outcome measure(s): Leptin and satiety. Results: Estradiol suppression significantly decreased leptin and increased lipid profiles (FDR-adjusted p≤0.05). Sleep fragmentation significantly increased heart rate (FDR-adjusted p=0.002) and trended to increase fasting glucose (FDR-adjusted p=0.08). Estradiol suppression and sleep fragmentation worsened individual cardiometabolic outcomes by (median, IQR) 4.0% (1.5%, 6.3%) from normalized baseline values. Sleep fragmentation worsened a composite cardiometabolic index derived from individual clinical cardiometabolic measures by an additional 103% over estradiol suppression alone. Conclusions: Independent of aging, there are significant adverse changes in cardiometabolic health induced by core components of the transition to postmenopause, including novel effects of sleep fragmentation, a modifiable target.Publication Predictive modeling of long-term improvement in occlusion outcomes following Woven EndoBridge treatment of cerebral aneurysms: A machine learning approach(2025-11-03)BackgroundThe Woven EndoBridge (WEB) device represents an innovative solution for cerebral aneurysm occlusion, particularly for challenging wide-neck bifurcation aneurysms. However, factors affecting sustained occlusion remain poorly understood. We utilized machine learning to attempt to identify predictors of favorable long-term outcomes following WEB treatment.MethodsIn this multicenter retrospective study, we collected patient demographics, aneurysm characteristics, procedural details, and clinical outcomes. The primary endpoint was improvement in occlusion status, defined as maintained Raymond-Roy Occlusion Classification (RROC) grade 1, or improvement from grade 2 to 1, or from grade 3 to either 2 or 1 on final angiographic follow up. The dataset was split into training (75%) and validation (25%) sets. The CatBoost algorithm was selected based on performance metrics, with Shapley Additive exPlanations (SHAP) values calculated to determine feature importance. Furthermore, a multivariable binomial logistic regression model was performed to validate machine learning findings.ResultsAmong 720 aneurysms from 36 hospitals, 84% showed improvement in occlusion at follow up. Both machine learning and multivariable logistic regression identified aneurysm height as the most consistent correlate of nonimprovement (odds ratio (OR) 0.90 per mm, p = 0.022). In the CatBoost model, the highest-ranking features by SHAP included aneurysm height, patient age, treatment acuity, ACom location, WEB-SLS device, bifurcation anatomy, aneurysm multiplicity, baseline modified Rankin Scale, access route, and partial thrombosis.ConclusionsMachine-learning and regression analyses identified consistent predictors of occlusion improvement after WEB treatment, with aneurysm height most strongly linked to nonimprovement. These insights may guide patient selection and follow up. Findings require cautious interpretation and external validation in larger cohorts.