Smartphone-based Pharmacovigilance Platform using AI for Adverse Event Detection

Thesis Abstract

The rapid proliferation of mobile health technologies presents an opportunity to strengthen pharmacovigilance by enabling real-time, user-reported adverse event (AE) data capture and analysis. However, existing systems are fragmented, suffer from underreporting, and lack scalable AI-driven mechanisms to classify and prioritize signals from heterogeneous data sources. This study develops and evaluates a smartphone-based pharmacovigilance platform that leverages artificial intelligence to detect and triage AEs from patient-reported outcomes, pharmacy records, and clinician inputs, thereby enhancing signal detection while maintaining user privacy and data integrity. The aim is to design, implement, and validate an integrated platform that (i) collects structured and unstructured AE reports through a mobile app, (ii) preprocesses and harmonizes data from multiple sources, (iii) applies natural language processing (NLP) and machine learning (ML) models to classify AEs and assess causality, and (iv) generates actionable safety signals for pharmacovigilance teams. Specific objectives include (1) to develop a user-centric data capture interface with guided reporting and multilingual support; (2) to implement NLP-based extraction of AE mentions from free-text narratives and integrate structured dosage and timing data; (3) to train and evaluate ML classifiers (random forest, gradient boosting, and transformer-based models) for AE categorization, severity assessment, and plausibility using a labeled dataset of 10,000 anonymized reports; (4) to construct a Bayesian network model for probabilistic causality assessment and signal prioritization; (5) to implement privacy-preserving analytics and secure data transmission compliant with GDPR and local regulatory frameworks; (6) to test platform performance in a 6-month pilot involving 2,000 volunteer participants and 15 collaborating healthcare facilities; (7) to compare platform-generated signals with conventional pharmacovigilance databases (e.g., FAERS) using concordance metrics and time-to-signal analysis. Methodologically, the study adopts a mixed-methods approach within a developmental- evaluative framework. A constructive design research method guides the iterative development of the platform, combining quantitative model evaluation with qualitative user feedback. The population comprises adult patients on prescription medications, with a sample of 2,000 app users and 15 clinicians/research pharmacists for mixed-method validation. Data collection instruments include the mobile AE reporting app, an optional clinician dashboard, structured questionnaires for usability (System Usability Scale) and trust in AI, and semi-structured interviews with clinicians to explore signal workflows. Validity and reliability are established through cross-validation of ML models (5-fold CV), external validation using a held-out dataset of 2,500 reports, and measurement invariance checks across languages. Data analysis methods encompass descriptive statistics, ML performance metrics (precision, recall, F1-score, AUROC), calibration plots for probabilistic outputs, thematic analysis for qualitative interviews, and time-to-signal analyses comparing platform versus standard reporting. Expected findings indicate that NLP-enabled extraction improves AE capture completeness by 28% and that transformer-based classifiers achieve AUROC of 0.92 for AE categorization and 0.89 for causality plausibility. The Bayesian network is anticipated to provide robust signal prioritization with a >70% concordance to expert-determined signals and a 25–40% reduction in time-to-signal relative to conventional pharmacovigilance processes. Privacy-preserving techniques (pseudonymization, on-device inference, and federated learning options) are expected to maintain data utility while meeting regulatory requirements. The study also anticipates favorable user experiences (SUS scores > 70) and high clinician satisfaction with the signal workflow. The contribution to knowledge includes (i) a scalable, AI-driven framework for interoperable pharmacovigilance using patient-generated health data, (ii) an empirically validated blend of NLP, ML, and probabilistic causality for AE detection and prioritization, (iii) a practical model for privacy-preserving, cross-site pharmacovigilance analytics, and (iv) evidence on the integration of mobile reporting into regulatory-grade signal management workflows. The main conclusion is that smartphone-based pharmacovigilance platforms leveraging AI can significantly enhance AE reporting quality, shorten signal detection timelines, and improve patient safety when designed with rigorous validation and robust governance. Recommendations emphasize expanding multilingual support, integrating with electronic health records to enrich context, adopting federated learning to balance data access with privacy, and conducting longitudinal impact assessments across diverse healthcare systems to generalize the model's applicability.