Interoperable Medical Device Analytics in Microservice Architectures: Development of a Hybrid Electrochemical and Gaussian Process Regression Framework for Lithium Carbon Monoflouride Battery Prognostics

Abstract

Reliable battery prognostics for active implantable medical devices (IMDs), such as cardiac pacemakers and neurostimulators, are critical for safeguarding patient health and avoiding premature or delayed surgical replacements. However, primary lithium carbon monofluoride (Li/CFx) cells exhibit an exceptionally flat discharge plateau where the voltage gradient is minimal, making state-of-charge (SoC) estimation notoriously difficult. Furthermore, clinical telemetry data collected during routine hospital follow-ups is inherently sparse and irregular, causing traditional mechanistic tracking models to accumulate drift and pure data-driven estimators to fail. To address these challenges, this thesis proposes an interoperable software architecture and hybrid prognostic framework deployed as a decoupled, stateless microservice. Exposing asynchronous RESTful endpoints via FastAPI, the service integrates Pydantic data models for schema validation at the network boundary and utilizes HL7 FHIR standards to support interoperable data exchange across clinical hospital systems. The core analytical engine couples a physical second-order equivalent circuit model (ECM) prior with a Gaussian Process Regression (GPR) statistical corrector to dynamically compensate for parameter mismatch and patient-specific load profiles. Programmatic thermodynamic guardrails are implemented to strictly enforce monotonic capacity depletion. A containerized simulation environment orchestrated via Docker Compose was developed to validate the system. Across a retrospective cohort of n = 87 active devices, the bounded hybrid observer achieved a Mean Absolute Error (MAE) of 1.346% and a Root Mean Squared Error (RMSE) of 5.654%, outperforming mechanistic (4.611%) and data-driven (6.370%) baselines. In addition, automated stress-testing verified that the centralized exception boundary safely isolated 100% of injected network and data anomalies. Latency benchmarks demonstrated processing execution scaling from 0.382 s (N = 1) to 2.411 s (N = 1000) on a local container mesh. This demonstrates that the microservice architecture successfully balances robust software engineering with high-accuracy algorithmic execution, underscoring its potential utility as an analytical foundation for safety-critical clinical monitoring environments pending future in-vivo and clinical validation.