Explainable Machine Learning for Classifying Documented Death Outcomes in FDA Adverse-Event Reports: A Multi-Quarter Exploratory Study and Pharmacovigilance Framework
Year: 2026
Type: preprint
Abstract: Objective To describe an explainable artificial intelligence pharmacovigilance framework and evaluate whether machine-learning models can classify FDA spontaneous adverse-event reports that document a death outcome, while clearly distinguishing retrospective report classification from prospective patient-level prediction. Methods We analyzed official FDA AEMS/FAERS ASCII quarterly extracts from 2020 Q1 through 2025 Q1 (21 quarters). Reports without a recorded outcome-table entry were excluded. A report was labelled positive when at least one OUTC row recorded the DE (death) code; other reports with recorded outcomes formed the comparison class. We excluded case identifiers listed in the supplied quarterly deletion files and retained one report version per CASEID using the highest CASEVERSION, then FDA receipt date and PRIMARYID as tie-breakers. A disk-backed SQLite workflow resolved versions across quarters. A bounded sample was drawn within chronological fit (2020–June 2023), calibration (July–December 2023), and test (2024 Q1–2025 Q1) windows. Logistic regression and histogram-based gradient boosting were fitted, sigmoid-calibrated on the separate calibration sample, and assessed on the held-out test sample. The operating threshold for each model was selected by maximizing F1 on the calibration sample. Subgroup metrics and TreeSHAP explanations were evaluated on sampled test records. Results After outcome filtering, deletion-list exclusion, case-version resolution and removal of records without usable FDA receipt dates, the analysis cohort contained 4,534,679 report versions/cases, including 667,062 reports with a recorded DE outcome and 3,867,617 reports with other recorded outcomes. The model sample included 250,000 fit reports, 75,000 calibration reports and 150,000 chronological test reports. In the test sample, gradient boosting achieved ROC-AUC 0.722 and average precision 0.343, compared with 0.657 and 0.247 for logistic regression. At thresholds selected on the calibration sample to maximize F1, gradient boosting achieved death-class precision 0.304, recall 0.469, F1 0.369 and balanced accuracy 0.645; its Brier score was 0.110. SHAP summaries ranked reaction-row count, age, sex, reporter type, occupation code and country-related fields among the leading model features. These are report-level associations and may reflect reporting and documentation processes. Conclusion The multi-quarter analysis demonstrates that a carefully defined retrospective report-classification task can be evaluated reproducibly across FDA quarterly extracts. It does not establish prospective prediction of ADRs or patient deaths, drug causality, or clinical utility. Same-report reaction and drug fields may be recorded after or alongside the outcome, and the study used a sampled test set rather than a full-census holdout. Independent validation and a prospectively timed feature design would be required before any early-warning claim or clinical deployment.
Source: Research Square
Authors: Nitin Triphati¹, Prashant Kumar Pandey²
Language: English
Subfield: Toxicology
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7236669926
