📄 medicine

Resting electrocardiographic and pulse-wave deep learningidentifies exaggerated exercise blood pressure response in elite athletes

This study demonstrates that deep learning models combining resting electrocardiograms and pulse waves can effectively identify an exaggerated exercise blood pressure response in elite athletes, offering a potential non-exercise screening tool for cardiovascular risk.

Original authors: Dominic Eckerle, Nils Gumpfer, Michael Guckert, Pascal Bauer, Jennifer Hannig

Published 2026-07-31

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Dominic Eckerle, Nils Gumpfer, Michael Guckert, Pascal Bauer, Jennifer Hannig

Original paper licensed under CC BY 4.0 (https://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

Imagine your body as a high-performance race car. When you're just sitting in the garage (resting), the engine idles quietly. But when you hit the track and floor the gas pedal (exercise), the engine roars, and the pressure in the fuel lines spikes. Most cars handle this surge smoothly. However, some cars have a quirk where the pressure shoots up way too high, too fast, just because of how the engine and pipes are built. In the human body, this is called an "exaggerated blood pressure response." It's like a warning light that flickers on, suggesting the car might be prone to trouble later, even if it looks fine right now.

Doctors have long known that checking this pressure surge during a workout is a great way to spot future heart trouble. But getting that data is like sending the car to a professional racing simulator: it takes time, expensive equipment, and a lot of effort. What if you could tell if a car is prone to these pressure spikes just by listening to the engine idle or feeling the vibration of the chassis while it's parked? That's the big question scientists are asking: Can we predict how a body reacts to a hard workout just by looking at its "resting signals"? This is the territory of deep learning, where computers act like super-smart detectives, scanning tiny patterns in heartbeats and pulse waves that human eyes might miss, hoping to find a shortcut to keeping athletes safe.

The Paper's Story: Can a Computer Guess a Workout's Pressure Spike from a Nap?

In this study, a team of researchers played detective with a group of elite athletes—mostly professional handball, basketball, and football players. They wanted to see if a computer could look at two simple, resting signals and guess whether an athlete would have a dangerous blood pressure spike during a workout.

The two signals were:

The ECG (Electrocardiogram): Think of this as the electrical spark plug map of the heart. It shows how the heart's electrical system fires to make the muscle squeeze.
The RPW (Radial Pulse Wave): This is the physical "thump-thump" wave of blood traveling through the arteries in the wrist. It's like feeling the ripple in a hose when water rushes through it.

The athletes had to do the hard part first: a grueling bike test where their blood pressure was measured at every step of the climb. The researchers used this data to label the athletes. If an athlete's blood pressure rose too sharply for the amount of effort they put in (specifically, a slope greater than 6.2 mmHg per MET), they were tagged as having an "exaggerated response." About 27% of the 334 check-ups in the study fell into this "high-risk" category.

Then came the fun part: They fed the resting ECG and pulse wave data into a "deep learning" brain—a type of computer program called a Convolutional Neural Network (CNN). Imagine this AI as a student who has never seen the bike test results. It was only shown the resting heartbeats and wrist pulses and asked, "Based on this, will this person's blood pressure go crazy when they exercise?"

What Did the Computer Find?

The results were a mix of "pretty good" and "not quite there yet."

The Heartbeat Won: The computer was much better at guessing the answer using just the ECG (the electrical spark map). It got it right about 77% of the time on the test group. The AI seemed to focus on specific parts of the heartbeat, particularly the big "QRS" spikes (the main electrical explosion that makes the heart beat), suggesting that the shape of the heart's electrical signal holds clues about how it handles stress.
The Pulse Wave Was a Maybe: The computer tried using just the wrist pulse wave (RPW), but it was less sure, getting it right about 66% of the time. It seemed to pay attention to the very beginning of the pulse wave, the moment the blood first shoots out.
The Power of Two: When the researchers combined both signals—letting the computer look at the electrical map and the physical pulse wave together—it got slightly better, reaching about 79% accuracy. It's like having two detectives instead of one; they didn't completely solve the case together, but they did catch a few more clues than either could alone.

The Catch: It's a Clue, Not a Crystal Ball

Here is the most important part of the story: The authors are very careful not to say they have found a magic cure. They suggest that this method could be a helpful screening tool, like a smoke detector that goes off before a fire starts. If the computer flags an athlete as "risky" while they are just resting, doctors might decide to give them a more thorough, expensive bike test to be sure.

However, the paper explicitly rules out the idea that this resting test can replace the real workout test. The computer isn't perfect yet; it misses some cases, and the confidence intervals (the range of how sure we are) are quite wide. The researchers admit that because they only looked at young, mostly male professional athletes, we don't know if this trick works for older people, women, or regular gym-goers.

The Bottom Line

This study suggests that the resting heart and pulse waves of elite athletes do contain hidden "secret codes" that hint at how their blood pressure will react to exercise. The computer learned to read these codes, especially from the electrical heartbeat, with a decent level of success. But right now, this is just a promising "proof of concept." It's a cool new tool in the toolbox, but until it's tested on more diverse groups and proven to predict future heart problems, it's not ready to take the place of the hard work of a real exercise test. The AI is a helpful sidekick, but the doctor still needs to see the athlete run the race to be absolutely sure.

Technical Summary: Resting Electrocardiographic and Pulse-Wave Deep Learning Identifies Exaggerated Exercise Blood Pressure Response in Elite Athletes

Problem Statement
An exaggerated systolic blood pressure response to exercise (eBPR) is a known risk factor for future arterial hypertension and adverse cardiovascular outcomes. However, eBPR is currently defined as an exercise-derived phenotype with heterogeneous definitions, often relying on workload-indexed thresholds such as a systolic blood pressure-to-metabolic equivalent (SBP/MET) slope greater than 6.2 mmHg/MET. Identifying eBPR typically requires resource-intensive, standardized exercise testing (e.g., cycle ergometry), which limits its feasibility for large-scale or repeated screening, particularly in athletic populations. While resting cardiovascular biosignals, such as 12-lead electrocardiograms (ECG) and radial pulse waves (RPW), contain information regarding vascular and cardiac properties, it remains unknown whether latent patterns within these resting time-series signals can identify individuals with eBPR without the need for exercise testing. Previous studies have largely relied on predefined signal-derived parameters rather than analyzing raw time-series data directly.

Methodology
This retrospective study utilized a cohort of 197 predominantly male professional athletes (aged 18–39) who underwent 334 pre-season cardiovascular screening examinations between 2017 and 2023. The dataset included resting 12-lead ECGs, radial pulse waves (RPW) obtained via oscillometric analysis, and standardized cycle ergometry results used to define the ground truth.

Data Preprocessing:
- ECG: Raw 5-second, 12-lead recordings (500 Hz) were segmented into overlapping 4-second windows. This resulted in 1,670 ECG windows (5 windows per recording).
- RPW: Oscillometric recordings were filtered for quality; the first four valid recordings per examination were retained and resampled to a uniform length of 1,000 data points via linear interpolation.
- Labels: Examinations were labeled as "eBPR" if the SBP/MET slope exceeded 6.2 mmHg/MET (92 cases, 27.5%) and "non-eBPR" otherwise. Labels were assigned at the examination level, not the participant level.
Model Architecture:
- Separate one-dimensional Convolutional Neural Networks (CNNs) were developed for each modality.
- RPW Model: Input shape (1000 × 1). Architecture included six convolutional layers (64 filters each) with ELU activation and max-pooling, followed by global average pooling and three fully connected layers (128 units, 0.3 dropout).
- ECG Model: Input shape (2000 × 12). Architecture included seven convolutional layers (64 filters each) with ELU activation and max-pooling, followed by global average pooling and three fully connected layers (128 units, 0.5 dropout).
- Hyperparameters were optimized via grid search on the development set, selecting models based on validation AUC and Matthews Correlation Coefficient (MCC).
Training and Validation Strategy:
- To prevent information leakage, data splitting was performed at the subject level.
- The dataset was split into an independent hold-out test set (20% of subjects) and a development set (80%).
- The development set underwent group-stratified five-fold cross-validation, ensuring all data from a single participant remained within the same fold.
- Ensemble Prediction: Predictions from the five cross-validation models were aggregated using soft voting (arithmetic averaging) at the examination level.
- Late Fusion: A combined RPW–ECG model was created by averaging the ensemble probabilities of the modality-specific models at the decision level, rather than fusing features at the input level.
Explainability:
- Post-hoc Explainable AI (XAI) was applied using Layer-wise Relevance Propagation with the $\epsilon$ -rule and SIGN transformation (LRP- $\epsilon$ SIGN) to generate global explainability maps highlighting signal regions contributing to predictions.

Key Results

Single-Modality Performance:
- The ECG-only model demonstrated the strongest single-modality performance, achieving a mean test AUC of 0.77 (95% CI 0.62–0.90) with a sensitivity of 0.72 and specificity of 0.73.
- The RPW-only model achieved a lower mean test AUC of 0.66 (95% CI 0.50–0.80), with a sensitivity of 0.56 and specificity of 0.58.
Combined Model Performance:
- The late-fusion RPW–ECG model achieved the highest overall performance with a test AUC of 0.79 (95% CI 0.65–0.90).
- However, the improvement over the ECG-only model was marginal. At the default probability threshold of 0.5, the combined model showed a slightly lower sensitivity (0.67) compared to the ECG-only model (0.72), while specificity remained similar (0.73).
Explainability Findings:
- RPW: Relevance scores were concentrated in the early systolic phase of the pulse waveform, specifically around the initial pressure upstroke.
- ECG: Relevance was primarily focused on the QRS complexes and adjacent signal segments across multiple leads.

Significance and Claims
The authors position this study as a proof-of-concept demonstrating that resting cardiovascular biosignals contain latent information associated with the eBPR phenotype in elite athletes. The study claims that:

Feasibility: Deep learning models trained on raw resting ECG and RPW time series can identify eBPR with moderate accuracy, suggesting that resting signals encode physiological signatures related to exercise blood pressure regulation.
Modality Superiority: In this specific cohort, ECG signals provided the primary discriminative power, while RPW offered only limited incremental value when combined via late fusion.
Clinical Utility (Cautious): The authors explicitly state that these models cannot diagnose eBPR in the strict physiological sense, as eBPR is an exercise-derived phenomenon. Instead, the models estimate the probability that an examination would be classified as eBPR during standardized exercise testing.
Future Application: If externally validated, such a model could serve as a screening tool to prioritize athletes for more detailed exercise BP assessments, ambulatory monitoring, or longitudinal follow-up, potentially acting as a complement to, but not a replacement for, standardized exercise testing.

Limitations Acknowledged
The paper notes several constraints: the cohort is homogeneous (predominantly male, professional athletes), limiting generalizability to women, older individuals, or recreational populations; the dataset size is relatively small for deep learning; and the definition of eBPR relies on a specific SBP/MET threshold that lacks universal consensus. Furthermore, the late-fusion approach did not capture feature-level interactions between modalities.

Technical Summary: Resting Electrocardiographic and Pulse-Wave Deep Learning Identifies Exaggerated Exercise Blood Pressure Response in Elite Athletes

More like this