Neural Network Based Simulation of Diagnostic Vulnerabilities in Telemedicine Using Synthetic Data: A Framework for Identifying Context-Driven Blind Spots

Abstract

Although telemedicine has made healthcare more accessible, it also poses diagnostic risks when crucial background data, like prescription history and procedure details, is absent. We created a simulation framework that employs neural networks as stand-ins for clinical reasoning in order to examine these weaknesses. The Synthea® platform was used to create a dataset of 100,000 simulated patient encounters, which were then examined in two different scenarios: telemedicine with limited input and traditional in-person care with complete clinical context. On both input types, four neural models: MLP, Deep MLP, TabNet, and a custom PyTorch network were trained independently. F1 score degradation, prediction confidence, consistency across conditions, and confidence-based escalation were used to assess performance. With F1 score drops as high as 0.581, the results demonstrated significant diagnostic degradation in telemedicine settings, especially for conditions related to metabolism, cancer, and allergies or immunology. Additionally, stability and confidence declined, suggesting increased uncertainty. Uncertain cases were successfully identified by confidence-based escalation techniques, indicating a helpful safety measure. Overall, the results demonstrate that when critical clinical context is missing, telemedicine poses quantifiable diagnostic risks. The suggested framework offers a methodical approach to pinpoint condition-specific vulnerabilities and direct safer remote care procedures.

Ähnliche Arbeiten