Development and validation of a digital biomarker for peripheral artery disease

Figures

• 

  Fig. 1. Multi-database study cohort of toe photoplethysmography (T-PPG) measurements. Signals were removed due to missing blood pressure (BP) measurements, incompressible ankle-brachial index (ABI) values, or due to poor four-second T-PPG recordings. Note that 83 patients were shared across each database after exclusion. Excluded signals did not bias the dataset toward certain patient groups or peripheral artery disease (PAD) severity levels, see Supplemental Fig. 1 for details on excluded signals.

• 

  Fig. 2. Photoplethysmography (PPG) morphology features are correlated with ankle brachial index (ABI). Data points are color-coordinated for peripheral artery disease (PAD) disease level, consistent with the thresholds in Table 1, where green is healthy, borderline PAD is blue, likely PAD is pink, and severe PAD is red. a PPG feature of normalized width at half amplitude is statistically correlated with ABI (a1) (R2 = 0.297, slope = −2.319, p < 0.0001). A sample healthy pulse (a2) is compared to a sample PAD pulse (a3), where the normalized width at half amplitude is highlighted on each in blue. b PPG feature of maximum rising slope is statistically correlated with ABI (b1) (R2 = 0.294, slope = 2.308, p < 0.0001). b2 A sample healthy pulse (green) is compared to a sample PAD pulse (red), where the maximum rising slope is highlighted on each with the blue dotted line.

• 

  Fig. 3. Support vector machine (SVM) performance on clinical, photoplethysmography (PPG) waveform, and combinations of clinical and waveform features. Area under the curve (AUC) performance for SVM classifier over 10 folds. Each color denotes a different combination of clinical and waveform features. Demographics alone (including age, gender, race, and ethnicity) is shown in red. Demographics and comorbidities is shown in orange. Demographics, comorbidities, and smoking status is shown in green. PPG alone is shown in blue. PPG and smoking status is shown in pink. PPG, demographics, comorbidities, and smoking status combined is shown in black. a All models with PPG waveform features outperform clinical-only models in the receiver operating characteristic (ROC) curve. b All models with PPG waveform features outperform clinical-only models in the precision-recall (PR) curve. c ROC-AUC and PR-AUC results with 95% confidence interval (CI) for all combinations of PPG waveform, demographics, comorbidities, and smoking status features.

• 

  Fig. 4. Feature importance of photoplethysmography (PPG) waveform and clinical features. Mutual information scores of PPG waveform (blue) and clinical (orange) features. All waveform features outperform all clinical features in importance. Just the top 15 and bottom 15 of waveform features are shown for clarity.

• 

  Fig. 5. Enhanced photoplethysmography (PPG) model area under the curve (AUC) performance across cohort. a Receiver operating characteristic (ROC) curves for each subgroup compared to the overall model performance. Solid colored lines distinguish subgroups: Male (red), Female (crimson), Hispanic (dark green), Non-Hispanic (lime green), White (turquoise), Black (cyan), Asian (sky blue), Other/Mixed Race (dark blue), CAD Present (dark purple), Diabetes Present (pink), ESRD Present (hot pink) against Overall (black). b ROC-AUC with 95% confidence interval (CI) and two-sided p-values reported for each sex, ethnicity, race, and highlighted comorbidity subgroup. Note that subgroup performance metrics were not reported for ‘Unknown’ Sex due to insufficient sample size (N = 4). Abbreviations: CAD, coronary artery disease; ESRD, end-stage renal disease.