⚡ electrical engineering

CAP: Towards PPG Universal Representation Learning with Patient-level Supervision

This paper introduces CAP, a novel framework that leverages a large-scale paired PPG-EHR dataset and patient-level supervision to align photoplethysmography signals with clinical semantics, significantly improving representation learning and performance across diverse downstream health monitoring tasks.

Original authors: Chenyang He, Xinyi Shao, Shun Huang, Bosong Huang, Daoqiang Zhang, Ming Jing, Cheng Ding

Published 2026-06-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Chenyang He, Xinyi Shao, Shun Huang, Bosong Huang, Daoqiang Zhang, Ming Jing, Cheng Ding

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Teaching a Doctor's Ear to Listen to the Whole Story

Imagine you are trying to teach a computer to understand the human body by listening to a PPG signal. A PPG signal is a light-based sensor (like the one on your smartwatch) that measures your blood flow. It looks like a wavy line on a graph.

For a long time, computers have been trained to look at these waves in short, 30-second snapshots. They try to guess your heart rate or blood pressure just by looking at the shape of the wave in that tiny moment.

The Problem: The authors argue this is like trying to understand a person's entire life story by reading a single sentence they wrote on a napkin. You might get the grammar right, but you miss the context. Two people might have identical heartbeats on a napkin, but one is a healthy athlete and the other is in the middle of a heart attack. The computer, looking only at the "napkin" (the short wave), can't tell the difference.

The Solution: CAP (Clinical Anchored Pretraining)

The authors created a new system called CAP. Think of CAP as a medical student who doesn't just listen to the heartbeat; they also read the patient's entire medical history file before making a diagnosis.

Here is how they built it, step-by-step:

1. The New Library (The Dataset)

To teach the computer, they built a massive library of paired information.

The Old Way: Just the heartbeat wave.
The CAP Way: They took long recordings of heartbeats and paired them with the patient's Electronic Health Records (EHR).
The Magic Trick: They used an AI language model to read messy, fragmented doctor's notes and turn them into a clean, organized "patient story." Now, every heartbeat wave is paired with a summary of that patient's health history, past illnesses, and current state.

2. The Training Method (The "Anchoring")

They trained the computer using three specific exercises, like a rigorous gym routine for an AI:

Exercise A: The "Fill-in-the-Blanks" Game (Morphological Reconstruction)
Imagine covering parts of a heartbeat wave with a black marker and asking the computer to draw the missing parts. This teaches the AI the basic "grammar" of a heartbeat (where the peaks and valleys should be).
Exercise B: The "Steady Hand" Challenge (Physiological Stability)
They added static noise (like radio interference) to the waves and asked the computer to ignore the noise and find the true signal. This teaches the AI to be robust against bad sensors or movement.
Exercise C: The "Context Anchor" (Clinical Anchoring)
This is the most important part. They showed the computer a heartbeat wave and its corresponding "patient story" (EHR). They forced the computer to learn that this specific wave shape belongs to this specific health story.
- Analogy: It's like teaching a child that a picture of a crying baby isn't just "a baby crying," but specifically "a baby who is hungry" or "a baby who is sick," based on the context provided. The heartbeat is "anchored" to the patient's real-life health condition.

3. The Final Exam (Downstream Tasks)

After training, they tested the AI on four different real-world tasks:

Detecting Atrial Fibrillation (an irregular heartbeat).
Estimating Heart Rate.
Estimating Respiratory Rate (breathing speed).
Estimating Blood Pressure.

The Results: Why It Won

The paper claims that CAP was significantly better than all previous methods.

The "Breathing" Breakthrough: The biggest win was in predicting how fast a person is breathing. CAP improved accuracy by 87.6% compared to the previous best method.
- Why? Breathing is a slow, long-term process. Short snapshots miss the pattern. Because CAP looked at the "whole story" (long-term context) and the patient's history, it could see the breathing pattern much better.
Overall Success: Across all four tasks, CAP was about 26.7% better on average than the competition.

How It Works in Practice (The "Gated" Brain)

When the AI makes a prediction, it uses a clever "gatekeeper" mechanism.

If the heartbeat signal is clear and high-quality, the AI focuses on the shape of the wave (the local details).
If the signal is messy, noisy, or shaky (like when you are running), the AI automatically shifts its focus to the patient's history and long-term context it learned during training.
Analogy: Imagine driving a car. If the road is clear, you look at the lane markings (the wave). If it starts raining and the road is slippery, you rely more on your GPS and knowledge of the route (the patient history) to stay safe.

Summary

The paper argues that to truly understand a heartbeat, you can't just look at the wave; you have to know who the wave belongs to. By teaching the AI to link heartbeats with patient medical histories, they created a system that is much smarter, more accurate, and better at handling messy, real-world data than anything that came before it.

Technical Summary: CAP – Towards PPG Universal Representation Learning with Patient-level Supervision

1. Problem Statement

Photoplethysmography (PPG) is a cornerstone of wearable health monitoring and clinical decision support, enabling applications such as heart rate monitoring, blood pressure estimation, and arrhythmia screening. However, existing approaches to Universal PPG Representation Learning face three critical limitations that hinder their generalization to complex clinical tasks and heterogeneous cohorts:

Semantic Sparsity and Bias: Current self-supervised learning (SSL) paradigms largely rely on signal-level objectives, such as geometric waveform similarity. This overlooks the fact that similar PPG morphologies do not necessarily imply identical physiological or pathological states, leading to a mismatch between learned representations and clinical reality.
Augmentation-Induced Semantic Corruption: Common data augmentation techniques (e.g., masking, temporal perturbations) used in SSL can inadvertently distort fine-grained structures carrying clinically relevant information, thereby constraining transferability to sensitive estimation tasks.
Limited Temporal-Scale Awareness: Most pretraining methods utilize short signal segments (e.g., 5–30 seconds), failing to capture long-term physiological dynamics and chronic conditions that are tightly coupled with a patient's broader clinical context.

The authors argue that PPG representation learning must shift from a signal-centric view to a patient-centric approach, incorporating long-horizon patient background information to anchor physiological representations in clinical semantics.

2. Methodology

2.1 Data Construction: Paired PPG-EHR Dataset

To support patient-centric learning, the authors constructed a large-scale multimodal dataset involving 2,279 patients. The dataset pairs long-term PPG signals with Electronic Health Records (EHR) distilled from fragmented medical histories.

Temporal Causality: Strict temporal alignment is enforced; for a PPG segment ending at time $T_{current}$ , the associated text includes only information available up to $T_{current}$ (excluding future discharge diagnoses or outcomes) to prevent look-ahead bias.
Hierarchical Alignment: Clinical text is organized into static context (demographics), cumulative context (past history), and concurrent context (proximal observations).
Distillation: A pretrained Language Model (PLM) distills raw clinical notes and diagnosis codes into structured EHR documents ( $e_k$ ) paired with 5-minute PPG windows ( $x_k$ ).
Quality Control: Signal Quality Indices (SQIs) are computed to filter out noisy segments, ensuring training on reliable physiological data.

2.2 The CAP Framework

The proposed Clinical Anchored Pretraining (CAP) framework bridges the gap between low-level physiological fluctuations and high-level clinical semantics through three core components:

A. Physical-Semantic Unified Patching (PSUP)

To address the granularity mismatch between long pretraining segments (5 minutes) and short downstream tasks (e.g., 30 seconds), CAP introduces PSUP. This mechanism dynamically resamples raw signals to a target frequency and slices them into non-overlapping patches. This ensures that the learned representation granularity is physically consistent and temporally aligned with diagnostic windows across different tasks.

B. Multi-Objective Pretraining

CAP optimizes three cooperative objectives during the pretraining stage:

Morphological Reconstruction Loss ( $L_{morph}$ ): Uses Masked Signal Modeling (MSM) to recover masked waveform details, forcing the encoder to learn the fundamental "grammar" of PPG (e.g., systolic peaks, dicrotic notches).
Physiological Stability Loss ( $L_{inv}$ ): A self-supervised contrastive objective that enforces feature consistency in the latent space under stochastic perturbations (e.g., dropout), ensuring robustness against sensor noise and baseline wander.
Clinical Anchored Loss ( $L_{clin}$ ): A cross-modal contrastive loss that aligns the aggregated global PPG representation with the structural EHR embedding derived from the PLM. This anchors physiological fluctuations to patient-level clinical semantics, mitigating representation drift.

C. Downstream Adaptation (Finetuning)

In the finetuning stage, CAP employs a dual-path architecture:

Local Morphology Encoder: Processes the raw PPG signal along with its first- and second-order temporal derivatives to capture local slope and curvature cues.
Clinical-Semantic Encoder: Reuses the pretrained PPG encoder to extract long-context clinical representations.
Gated Fusion: A gating mechanism adaptively balances the contribution of local morphological features and global clinical semantics based on the input, allowing the model to prioritize fine-grained details for high-quality signals or rely on robust long-context semantics when local features are noisy.

3. Key Contributions

Diagnosis of Current PPG SSL: The paper provides a systematic analysis of why existing signal-only SSL paradigms fail to achieve universal transfer, highlighting the lack of patient-level context.
Large-Scale PPG-EHR Dataset: The curation of a paired multimodal dataset that injects long-horizon patient background information into representation learning.
CAP Framework: The proposal of a pretraining framework that uses patient-level clinical semantics as an anchor, combined with Physical-Semantic Unified Patching to handle duration mismatches.
Empirical Validation: Comprehensive evaluation demonstrating that CAP learns clinically grounded, transferable representations across four diverse downstream tasks.

4. Experimental Results

The authors evaluated CAP on four heterogeneous downstream tasks using public benchmarks: Atrial Fibrillation (AF) detection, Heart Rate (HR) estimation, Respiratory Rate (RR) estimation, and Blood Pressure (BP) estimation.

AF Detection (Stanford Dataset): CAP achieved an F1-score of 0.883, outperforming strong baselines including GPT-PPG (0.841) and SiamQuality (0.770).
HR Estimation (WESAD Dataset): CAP achieved a Mean Absolute Error (MAE) of 3.99 (LOSO evaluation), surpassing the best baseline (GPT-PPG 1B) which scored 4.98.
RR Estimation (BIDMC Dataset): CAP demonstrated a massive improvement, achieving an MAE of 0.11 compared to the state-of-the-art baseline (GPT-PPG 1B) at 0.93. This represents a relative improvement of up to 87.6%.
BP Estimation (PulseDB): CAP achieved an MAE of 8.04, outperforming baselines like GPT-PPG (8.12) and SiamQuality (8.60).

Overall Performance: CAP delivered an average relative improvement of 26.7% across all four tasks compared to the strongest baselines.

5. Significance and Claims

The paper claims that CAP successfully shifts the paradigm of PPG representation learning from waveform fitting to modeling patient-level physiological states. By anchoring PPG representations to clinical semantics derived from EHR, the model achieves:

Enhanced Generalization: The ability to transfer effectively across diverse tasks and cohorts where signal-only models struggle.
Robustness: Improved performance in noisy conditions, as evidenced by the gating mechanism's ability to rely on long-context semantics when local signal quality degrades.
Interpretability: Visualizations (attention maps and UMAP) confirm that the model attends to clinically meaningful waveform regions and creates a more discriminative representation space for pathological states (e.g., clearer separation of AF-positive samples).

The authors conclude that incorporating long-term patient context and clinical semantics is a promising direction for developing universal foundation models for physiological signals, moving beyond the limitations of short-window, signal-level pattern matching.