Network Analysis and Machine Learning for Signal Detection and Prioritization Using Electronic Healthcare Records and Administrative Databases: A Proof of Concept in Drug-Induced Acute Myocardial Infarction

Background Safety signals for potential drug-induced adverse events (AEs) typically emerge from multiple data sources, primarily spontaneous reporting systems, despite known limitations. Increasingly, real-world data from sources such as electronic health records (EHRs) and administrative databases are leveraged for signal detection. Although network analysis has shown promise in mapping relationships between clinical attributes for signal detection in spontaneous reporting system databases, its application in real-world data from EHRs and administrative databases remains limited. Objective This study aimed to evaluate the performance of network analysis in detecting safety signals within Italian administrative databases, using drug-induced acute myocardial infarction (AMI) as a proof of concept. Methods We employed a case–crossover design to explore the association between drug exposure and AMI using the Healthcare Administrative Database of Mantova, Italy, from 2014 to 2018. Patients with their first AMI hospitalization were identified after a 365-day washout period to exclude prior hospitalizations. We constructed a network to analyse the relationships between prescribed drugs and diagnoses, represented as nodes, with undirected edges illustrating their interactions. For each patient with AMI, we identified all diagnoses and drugs recorded or redeemed within 365 days of the first AMI episode and generated various drug–diagnosis, drug–drug, and diagnosis–diagnosis pairs. We calculated the frequency of these pairs, and three types of edge weights quantified the strength of connections. We identified outlier drug–AMI pairs using a predictive score (F) based on frequency (C) and full edge weights (WF), with validation for known AMI associations. We prioritized signals using the F score, C of AMI, and WF, analysed through k-means clustering to identify patterns in the data. Results From 2014 to 2018, a total of 3918 patients had an AMI, with 4686 AMI diagnoses. Of those, 2866 had prescriptions in the previous year, totalling 498,591 prescriptions. A network analysis identified 2968 unique nodes, revealing 529,935 diagnosis–diagnosis connections, 235,380 drug–diagnosis connections, and 102,831 drug–drug connections. The median number of connections (C) was 404 (Q1–Q3: 194–671) for drug nodes and 380 (Q1–Q3: 216–664) for diagnosis nodes. The median WF was 11.8 (Q1–Q3: 9–14), and the median F score across pairs was 0.1 (Q1–Q3: 0.1–0.3). A total of 249 potential safety signals were detected, with 63.4% aligning with known AEs. Among the remaining signals, 80 were prioritized, and five emerged as the highest priority: terazosin, tamsulosin, allopurinol, esomeprazole, and omeprazole. Conclusions Overall, our novel method demonstrates that network analysis is a valuable tool for signal detection and prioritization in drug-induced AEs based on EHRs and administrative databases.