JMIR Publications

ABSTRACT

Background:

In many healthcare systems, governments can deploy medical equipment to remote regions; however, the persistent shortage of skilled professionals capable of interpreting complex diagnostic data continues to impede adequate care. This shortage often leads to misinterpretation, clinical errors, or preventable loss of life in underserved communities.

Objective:

To develop and evaluate a cloud-oriented smart diagnostic system that combines Long Short-Term Memory (LSTM) networks with the GPT2 model for the automated analysis of biomedical signals.

Methods:

The framework employs several PhysioNet datasets, such as the MIT-BIH Arrhythmia, PTB Diagnostic ECG, and Sleep-EDF databases, as key sources of physiological information. After pre-processing and signal conditioning, the LSTM model is used to analyze the data, extract temporal features, and generate probabilistic outputs. The results are then handled by GPT-2, which produces understandable, clinically significant explanations that illuminate otherwise unclear medical predictions.

Results:

Performance assessment on six data sets achieved accuracy as high as 94.7%, and AUC figures to 0.97. An expert evaluation conducted by three board-certified cardiologists confirmed the interpretations produced by GPT2, resulting in average ratings of 4.3 ± 0.4 (accuracy), 4.6 ± 0.3 (clarity), and 4.2 ± 0.5 (actionability).

Conclusions:

The system operates as a digital biomedical technician, allowing non-specialist health workers and clinicians to understand diagnostic results. The suggested framework could improve access to healthcare, dependability, and decision-making in areas with limited resources and remote locations.