Longitudinal artificial intelligence-based deep learning models for diagnosis and prediction of the future occurrence of polyneuropathy in diabetes and prediabetes

Objective

The objective of this study was to develop artificial intelligence-based deep learning models and assess their potential utility and accuracy in diagnosing and predicting the future occurrence of diabetic distal sensorimotor polyneuropathy (DSPN) among individuals with type 2 diabetes mellitus (T2DM) and prediabetes.

Methods

In 394 patients (T2DM=300, Prediabetes=94), we developed a DSPN diagnostic and predictive model using Random Forest (RF)-based variable selection techniques, specifically incorporating the combined capabilities of the Clinical Toronto Neuropathy Score (TCNS) and nerve conduction study (NCS) to identify relevant variables. These important variables were then integrated into a deep learning framework comprising Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks. To evaluate temporal predictive efficacy, patients were assessed at enrollment and one-year follow-up.

Results

RF-based variable selection identified key factors for diagnosing DSPN. Numbness scores, sensory test results (vibration), reflexes (knee, ankle), sural nerve attributes (sensory nerve action potential [SNAP] amplitude, nerve conduction velocity [NCV], latency), and peroneal/tibial motor NCV were candidate variables at baseline and over one year. Tibial compound motor action potential amplitudes were used for initial diagnosis, and ulnar SNAP amplitude for subsequent diagnoses. CNNs and LSTMs achieved impressive AUC values of 0.98 for DSPN diagnosis prediction, and 0.93 and 0.89 respectively for predicting the future occurrence of DSPN. RF techniques combined with two deep learning algorithms exhibited outstanding performance in diagnosing and predicting the future occurrence of DSPN. These algorithms have the potential to serve as surrogate measures, aiding clinicians in accurate diagnosis and future prediction of DSPN.