Beyond the Cuff: Deep Learning-Based Analysis of Handheld Doppler Signals to Detect Masked Peripheral Ischemia in Diabetic Patients with Normal Ankle-Brachial Index.

OBJECTIVE: The Ankle-Brachial Index (ABI) remains the primary screening tool for Peripheral Arterial Disease (PAD). However, its diagnostic performance is severely compromised in patients with diabetes mellitus due to medial arterial calcification (MAC), leading to falsely normal or elevated results. We aimed to evaluate the efficacy of a smartphone-based Deep Learning (DL) model analyzing handheld Doppler spectrograms to identify "masked" ischemia in symptomatic diabetic patients with pseudo-normal ABI. METHODS: In this prospective study, 120 participants were stratified into three cohorts: Group 1 (Masked Ischemia; n=40), symptomatic diabetic patients with false-normal ABI (0.90-1.30) but PAD confirmed via imaging; Group 2 (Overt PAD; n=40) with ABI <0.90; and Group 3 (Healthy Controls; n=40). Doppler audio signals were converted into Mel-Spectrograms. To account for intra-patient clustering, all diagnostic accuracy metrics and their 95% Confidence Intervals (CIs) were adjusted using Generalized Estimating Equations (GEE). To ensure robustness and eliminate data leakage, a Convolutional Neural Network (CNN) was trained using patient-level Leave-One-Patient-Out Cross-Validation (LOPOCV). Explainable AI (Grad-CAM) was utilized to visualize pathophysiologically relevant spectral features. RESULTS: As expected from the study design, ABI did not discriminate Group 1 from healthy controls (p=0.454). In contrast, accounting for intra-patient dependency via the GEE framework, the AI model successfully identified pathological flow in 37/40 patients in Group 1, achieving a cluster-adjusted sensitivity of 92.5% (95% CI: 80.1% - 97.4%) (p<0.001 vs. ABI). For the total cohort (n=480 signals), the model demonstrated a cluster-adjusted global accuracy of 95.8 %, a sensitivity of 93.7 %, and a specificity of 95.0 % . ROC analysis confirmed diagnostic superiority (AUC: 0.96 vs. 0.51; p<0.001). Notably, the AI-predicted risk score showed a strong positive correlation with duplex-derived Acceleration Time (AT) (r=0.78, p<0.001), validating its physiological relevance. CONCLUSIONS: Deep learning analysis of handheld Doppler signals demonstrated incremental diagnostic value in identifying peripheral ischemia where traditional pressure-based measurements face limitations. By shifting from mechanical pressure to flow-based hemodynamics, this point-of-care technology bridges a critical diagnostic gap and may reduce delayed diagnoses and subsequent limb loss in high-risk diabetic populations.