A Dual‐Modal Wearable PPG Smartwatch with AI‐Enhanced Correction for High‐Accuracy and Continuous AF Burden Assessment

FIGURE 1. AI‐Enhanced Dual‐Modal Smartwatch and Patch System for Accurate Atrial Fibrillation Burden Assessment. This figure illustrates the hardware, data processing workflow, and diagnostic performance of the dual‐modal wearable system integrating smartwatch PPG/ECG and patch‐based ECG with AI correction. (A) Exploded view of the patch components, including upper and lower shells, power button, battery, main PCB, and light PCB. (B) Exploded view of the smartwatch components, including screen, main PCB, sensor switch, battery, housing, sensor PCB, back cover, and electrodes. (C) Workflow showing simultaneous W‐PPG and W‐ECG signal acquisition. W‐ECG was used to correct W‐PPG via feature extraction and deep learning–based AF classification, and patch‐based ECG (P‐ECG) was used as the gold standard for evaluation. (D) Radar plot comparing model performance before (blue) and after (orange) AI correction. Metrics include AUC, SE, SP, PPV, NPV, ACC, and F1 score, demonstrating substantial improvement after correction. Abbreviations: AF: Atrial Fibrillation; W‐ECG: Watch‐based Electrocardiogram; W‐PPG: Watch‐based Photoplethysmography; P‐ECG: Patch‐based Electrocardiogram; AI: Artificial…

FIGURE 2. Smartwatch‐Based PPG Sensor and Touch‐Based ECG: Structural and Functional Overview. This figure illustrates the physical structure, optical working principle, ECG acquisition, and interpretation of systolic and diastolic components in the smartwatch‐based cardiovascular monitoring system. (A) Watch‐Based PPG Sensor: Physical Structure and Optical Path. Exploded view of the smartwatch PPG sensor, showing the LED light source, photodiode detector, and light propagation through the epidermis, dermis, and subcutaneous tissue to detect radial artery pulsation. (B) Optical Working Principle: Pulsatile Blood Volume Change. Schematic illustrating how changes in blood volume during systole and diastole affect reflected light intensity, forming the basis for PPG signal acquisition. (C) Touch‐Based Watch ECG. Illustration of single‐lead ECG acquisition via two dry electrodes (an upper/front electrode contacted by the opposite‐hand finger, and a back electrode in contact with the wearing wrist), activated by an ECG capacitive touch switch, and the resulting single‐lead W‐ECG waveform (P, Q, R, S, and T). (D) PPG Waveform Interpretation and Cardiac Conduction System. Conceptual diagram…

FIGURE 3. Smartwatch‐Based PPG Sensing and AI‐Based Atrial Fibrillation Detection Workflow. This figure illustrates the hardware configuration, optical signal acquisition, signal processing, and AI‐enabled AF classification of the smartwatch‐based PPG monitoring system. (A&B) Hardware Configuration. Top and side view of the smartwatch PPG sensor, showing the cover material (black PC+ABS), transparent PC lens, tri‐color LED transmitters, and photodiode detectors (VEMD8080) used for optical signal collection. (C) Signal Processing and AI Classification Workflow. Raw PPG signals undergo signal conditioning (filtering and noise removal), feature extraction (R‐R sinus rhythm vs atrial fibrillation), and deep learning classification to output AF or normal rhythm. (D) Diagnostic Outcome and Wearable Integration. Final classification output displayed on smartwatch interface, highlighting non‐invasive, real‐time, and continuous monitoring with AI‐based correction. Abbreviations: PPG: Photoplethysmography; AF: Atrial Fibrillation; PC: Polycarbonate; ABS: Acrylonitrile Butadiene Styrene.

FIGURE 4. Workflow for AI‑enhanced AF burden estimation from wearable signals. This figure outlines the comprehensive process of data acquisition, model implementation, feature matching, and label correction to enhance the accuracy of AF burden assessment using wearable devices. (A) Data input and management. Continuous photoplethysmography (W‐PPG) and single‐lead electrocardiography (W‐ECG) are recorded by the smartwatch. User‐initiated triggers, algorithm‐detected PPG events and scheduled prompts ensure synchronous acquisition of W‐ECG segments for subsequent calibration. (B) AI model implementation. Continuous W‐PPG data are segmented and processed by a parallel CNN‐LSTM model. The CNN branch extracts local waveform morphology and short‐term structural features, whereas the LSTM branch captures long‐range temporal dependencies and rhythm dynamics. The feature representations from the two branches are concatenated and passed through a fully connected classification layer to output a predicted rhythm label, either atrial fibrillation (AF) or non‐AF, for each segment.(C) AI Correction logic. Feature vectors derived from the W‐PPG segments are compared with template features extracted from…

FIGURE 5. Architecture of the Convolutional–recurrent Neural Network used for AI‐assisted Correction of Smartwatch PPG signals in AF Burden Assessment. The raw PPG and single‐lead ECG waveforms are simultaneously fed into a deep learning framework composed of a CNN module and an LSTM module. The CNN module acts as a hierarchical feature extractor: four parallel 1D convolutional layers capture multi‐scale morphological patterns, followed by 15 convolution blocks for deeper representation learning. Each convolution block consists of BN, Conv1d, MaxPool, and residual connections. AVGPool, the features are passed through fully connected (FC) layers. The LSTM module is designed to capture the temporal dependencies within the sequential signals. At each time step, the input sequence is processed by LSTM cells, and the hidden states are propagated across time. The outputs of the CNN and LSTM branches are concatenated and further passed through FC layers to produce the final classification result. Abbreviations: PPG, Photoplethysmography; ECG, Electrocardiogram; CNN, Convolutional neural network; LSTM, long short‐term memory network; Conv1d, One‐dimensional convolution; BN, Batch normalization;…

FIGURE 6. Feature Matching‐based Label Correction Framework for W‐PPG Segments. W‐PPG segments were divided into Cohort A and Cohort B according to the availability of concurrent W‐ECG recordings within the same acquisition window. Cohort A contained W‐PPG segments without paired W‐ECG and was used as the unlabeled set for calibration, whereas Cohort B contained synchronous W‐PPG segments with available W‐ECG and served as the template pool. Based on W‐ECG‐derived rhythm labels, Cohort B was further split into a positive template set (Cohort B‐P) and a negative template set (Cohort B‐N). For each segment in Cohort A, an intermediate feature representation, F (·), was extracted from the rhythm recognition model, and its similarity to the two template sets was quantified by nearest‐neighbor Euclidean distance. Using a predefined threshold (τ), samples were assigned to four subsets: SP, matched only to the positive template set; SN, matched only to the negative template set; SPN, matched to both template sets; and NS, matched to neither template set. Labels were then recalibrated according to the matching results: SP was reassigned as AF, SN as Non‐AF, whereas SPN and NS retained the original…

FIGURE 7. Correction Effects of W‐ECG–W‐PPG Template Similarity Matching. (A) When the W‐ECG classifier outputs a positive result, the circle centered at the P‐core denotes the region of W‐PPG templates that, according to similarity matching, can be re‐labeled as positive. The samples within this circle consist of false negatives (FN; green) and true positives (TP; blue). After calibration, FN are converted to TP, eliminating missed detections, while TP remain positive, leaving correct predictions unchanged. Overall, this procedure converts FN into TP and enhances classification accuracy. (B) When the W‐ECG classifier outputs a negative result, the circle centered at the N‐core represents the range of W‐PPG templates that can be corrected to the negative label. Samples inside this region include false positives (FP; grey), true positives (TP; blue) and true negatives (TN; orange). Following calibration, TN retain their label, FP are re‐classified as TN, and TP are re‐classified as false negatives (FN). Because the proportion of FP in this region is markedly larger than that of TN, the ratio of correct adjustments (FP→TN) exceeds that of incorrect adjustments (TP→FN), and the calibration…

FIGURE 8. Scatter and Bland‐Altman Plots for W‐PPG Correction by W‐ECG. Comparison of AF burden estimated from wearable PPG recordings and reference ECG measurements. (A) Correlation analysis before correction. (B) Correlation analysis after correction. In panels A and B, the x‐axis represents AF burden estimated by the wearable device (PPG AF burden, %) and the y‐axis represents AF burden measured by the reference ECG (ECG AF burden, %). (C) Bland–Altman analysis before correction. (D) Bland–Altman analysis after correction. In panels C and D, the x‐axis represents the mean AF burden (%) obtained from the wearable device and the reference ECG, and the y‐axis represents the difference in AF burden (%) between the wearable estimate and the reference ECG (PPG − ECG). (E) Comparison of absolute AF burden error before and after correction. All 1,054 participants were included in the analyses. Because multiple participants had similar or identical AF burden values, some data points overlap in the plots, resulting in fewer visually distinguishable points than the actual sample size. The figure was generated using enhanced visualization settings to improve the display of overlapping observations.…