Benjamin H. McMahon

Supplementary Data S1 from Polygenic Risk Scores for Prediction of Immune Checkpoint Inhibitor Thyroid Toxicity in Diverse Populations

Supplementary Methods, Tables, and Figures.

We present a data-driven framework to predict 15-year all-cause mortality using outpatient administrative records for 2.3 million Veterans in the largest integrated U.S. healthcare system. Rather than relying on predefined clinical phenotypes, we used the 1,000 most common outpatient medical codes from each of three data types/modalities – ICD-9 (Dx), Current Procedural Terminology (CPT), and prescription drugs (Rx), encoded as binary features. Using these features, we trained three machine learning (ML) (logistic regression with lasso, random forest, and a 3-layered feed-forward neural network) to predict 15-year mortality risk. The features were also mapped to variables for the widely used Charlson Comorbidity Index (CCI), Elixhauser, and Veterans Aging Cohort Study (VACS) indices, refitted for 15-year mortality prediction, for baseline comparison. All our models significantly outperformed the widely used CCI, Elixhauser, and VACS indices, with C-statistics ranging from 0.82 to 0.84 versus 0.739–0.804 for the baselines. Relative improvements in C-statistics of our approach over the baselines were consistent across different subgroups (age groups of<65 years, those 65+years, Blacks, Hispanics, etc.) Our approach enabled the identification of high-impact predictors with clinical grounding , without requiring hand-curated phenotypes. Cardiovascular diseases and mental health diagnoses/treatments emerged as leading long-term mortality indicators. Using unsupervised ML techniques including PCA and K-means clustering, we associated interpretable patterns and complex interactions between diagnoses and treatments, highlighting comorbidities, disease trajectories, and healthcare utilization patterns. The ability to achieve the predictive performance and algorithmically detect such relationships purely from outpatient data supports the scalability and broad applicability of our framework. This framework not only improves mortality risk stratification over existing clinical indices, but also enables better understanding of how medical codes, regardless of category, interact to predict long-term outcomes.

We present a data-driven framework to predict 15-year all-cause mortality using outpatient administrative records for 2.3 million Veterans in the largest integrated U.S. healthcare system. Rather than relying on predefined clinical phenotypes, we used the 1,000 most common outpatient medical codes from each of three data types/modalities – ICD-9 (Dx), Current Procedural Terminology (CPT), and prescription drugs (Rx), encoded as binary features. Using these features, we trained three machine learning (ML) algorithms (logistic regression with lasso, random forest, and a 3-layered feed-forward neural network) to predict 15-year mortality risk. The features were also mapped to variables for the widely used Charlson Comorbidity Index (CCI), Elixhauser, and Veterans Aging Cohort Study (VACS) indices, refitted for 15-year mortality prediction, for baseline comparison. All our models significantly outperformed the widely used CCI, Elixhauser, and VACS indices, with C-statistics ranging from 0.82 to 0.84 versus 0.739–0.804 for the baselines. Relative improvements in C-statistics of our approach over the baselines were consistent across different subgroups (age groups of <65 years, those 65+years, Blacks, Hispanics, etc.) Our approach enabled the identification of high-impact predictors with clinical grounding, without requiring hand-curated phenotypes. Cardiovascular diseases and mental health diagnoses/treatments emerged as leading long-term mortality indicators. Using unsupervised ML techniques including PCA and K-means clustering, we associated interpretable patterns and complex interactions between diagnoses and treatments, highlighting comorbidities, disease trajectories, and healthcare utilization patterns. The ability to achieve the predictive performance and algorithmically detect such relationships purely from outpatient data supports the scalability and broad applicability of our framework. This framework not only improves mortality risk stratification over existing clinical indices, but also enables better understanding of how medical codes, regardless of category, interact to predict long-term outcomes.

Purpose: Immune checkpoint inhibitor (ICI)–induced thyroiditis is a common immune-related adverse event (irAE) linked to improved survival. Polygenic risk scores (PRS) for autoimmune hypothyroidism predict thyroid irAEs in European-ancestry patients; performance in non-European populations is unclear. Experimental Design: In the Veterans Affairs Million Veteran Program (MVP; 2011–2023), we identified ICI-treated patients with germline genotyping and a chemotherapy-treated control cohort, excluding those with thyroid disease or prior thyroid-directed treatments. Harmonized ancestry and race/ethnicity (HARE) defined non-Hispanic White (NHW) and non-Hispanic Black (NHB) groups. Thyroid irAEs within 1 year were defined using laboratory criteria capturing both hyperthyroid and hypothyroid phases. We compared two PRSs: a published European-derived PRS and an updated PRS selected across multiple genome-wide association study (GWAS) sources and methods (including the MVP multiancestry GWAS) to maximize discrimination in African-ancestry individuals in a held-out test set. HARE-stratified multivariable Cox models estimated time to thyroiditis; a 6-month landmark analysis assessed overall survival. Results: The ICI cohort included 4,289 patients (3,473 NHW; 816 NHB). The baseline PRS was associated with thyroiditis in NHW [adjusted hazard ratio (aHR) per standard deviation, 1.33; 95% confidence interval (CI), 1.19–1.50) but not NHB patients or controls. The updated PRS improved risk stratification in NHW (aHR, 1.45; 95% CI, 1.28–1.63) and predicted thyroiditis in NHB patients (aHR, 1.48; 95% CI, 1.11–1.98) but not in controls. Thyroiditis within 6 months was associated with improved survival, but neither PRS was. Conclusions: Germline polygenic liability to hypothyroidism predicts ICI-induced thyroiditis in NHW and NHB patients when PRSs are selected via ancestry-stratified validation. Careful exploration of the dataset-method space is critical for equitable PRS development.

Background/Objectives: Prediction models are implemented frequently, yet, compared with other study designs, their incorporation of clinical measurements (CMs; i.e., vital signs and laboratory results) is rather underdeveloped. The purpose is to describe methods used and illustrate clinical utility in parameters systematically derived from CMs; as a case study, we use the risk of all-cause mortality following coronavirus disease 2019 (COVID-19) as the basis for prognosis. Methods: We identified cases through the Department of Veterans Affairs COVID-19 Shared Data Resource, utilizing data from the first visit until 14 days before testing positive. Thirteen parameters were derived from each of the 11 CMs, capturing departures from normality considering variability and time. The 143 candidate predictors were used to generate the main logistic regression model. The area under the receiver operating characteristic curve (AUROC) analysis was performed to assess discrimination between those who lived and died for subset and main regressions; for comparison, this was performed for an age-only model and the Charlson Comorbidity and Elixhauser Indices. Results: There were 329,491 patients. The main model’s AUROC (0.785 ± 0.002) was similar to the age-only model (0.783 ± 0.002; p > 0.05) and significantly greater than the comorbidity indices’ (range: 0.675 ± 0.002 to 0.729 ± 0.002; p < 0.001 each). Conclusions: The study found several parameters were significant determinants of mortality following COVID-19, highlighting the importance of a systematic approach for multivariate modeling to obtain informative insights into underlying pathophysiology. The main model outperforms common comorbidity indices as a summary metric for pre-existing conditions in this case study. If validated, this approach could revolutionize the way CMs are handled in multivariate models.

We present a stochastic differential equation model of suicidal progression in U.S. veterans, simulating transitions across mental health states under dynamic stress and covariate influence. Transition rates are modulated by an Ornstein-Uhlenbeck stress process and clinical features derived from retrospective case-control data. Simulations reveal profile-dependent tipping behavior, with risk-loaded individuals exhibiting persistent ideation and attempt states. Area-under-the-curve and phase plane analyses suggest early warning signals and support the use of individualized dynamical models for suicide risk assessment.

Background/Aims Predictive models of suicide risk have focused on predictors extracted from structured data found in electronic health records (EHR), with limited consideration of predisposing life events (LE) expressed in unstructured clinical text such as housing instability and marital troubles. Additionally, there has been limited work in large-scale analysis of natural language processing (NLP) derived predictors for suicide risk and integration of extracted LE into longitudinal models of suicide risk. This study aims to expand upon previous research, demonstrating how high-performance computing (HPC) and machine learning technologies such as language models (LM) can be used to annotate and integrate 8 LE across all Veterans Health Administration (VHA) unstructured clinical text data with enriched performance metrics. Materials/Methods VHA-wide clinical text from January 2000 to January 2022 were pre-processed and analyzed using HPC. Data-driven lexicon curation was performed for each LE by scaling a nearest-neighbor search over a precomputed index with LM embeddings. Data parallelism was applied to a rule-based annotator to extract LE, followed by random forest for improved positive predictive value (PPV). NLP results were analyzed and then integrated and compared to a baseline statistical model predicting risk for a combined outcome (suicide death, suicide attempt and overdose). Results First-time LE mentions, with a PPV of 0.8 or higher, showed a temporal correlation to suicide-related events (SRE) (suicide ideation, attempt and/or death). A significant increase of LE occurrences was observed starting 2.5 months prior to an SRE. Predictive models integrating NLP-derived LE show an improved AUC of 0.81 vs. a 0.79 obtained with the baseline and novel patient identification of up to 57%. Discussion Our analysis shows that: 1) performance metrics, specifically PPV, improved significantly from previous work and outperform related works; 2) the mentions of LE in the unstructured data increase as time to a SRE approaches; 3) LE identified from the notes in the weeks prior to a SRE were not associated with administrative bias caused by outreach; and 4) LE improved the AUC of predictive models and identified novel patients at risk for suicide. Conclusion The resulting person-period longitudinal data demonstrated that NLP-derived LE served as acute predictors for suicide-related events. NLP integration into predictive models may help improve clinician decision support. Future work is necessary to better define these LE.

Severe malarial anemia (SMA, Hb < 6.0 g/dL) is a leading cause of childhood morbidity and mortality in holoendemic Plasmodium falciparum transmission zones. This study explored the entire expressed human transcriptome in whole blood from 66 Kenyan children with non-SMA (Hb ≥ 6.0 g/dL, n = 41) and SMA (n = 25), focusing on host immune response networks. RNA-seq analysis revealed 6862 differentially expressed genes, with equally distributed up-and down-regulated genes, indicating a complex host immune response. Deconvolution analyses uncovered leukocytic immune profiles indicative of a diminished antigenic response, reduced immune priming, and polarization toward cellular repair in SMA. Weighted gene co-expression network analysis revealed that immune-regulated processes are central molecular distinctions between non-SMA and SMA. A top dysregulated immune response signaling network in SMA was the HSP60-HSP70-TLR2/4 signaling pathway, indicating altered pathogen recognition, innate immune activation, stress responses, and antigen recognition. Validation with high-throughput gene expression from a separate cohort of Kenyan children (n = 50) with varying severities of malarial anemia (n = 38 non-SMA and n = 12 SMA) confirmed the RNA-seq findings. Proteomic analyses in 35 children with matched transcript and protein abundance (n = 19 non-SMA and n = 16 SMA) confirmed dysregulation in the HSP60-HSP70-TLR2/4 signaling pathway. Additionally, glutamine transporter and glutamine synthetase genes were differentially expressed, indicating altered glutamine metabolism in SMA. This comprehensive analysis underscores complex immune dysregulation and novel pathogenic features in SMA.