Generalize Then Personalize: Wearable-Based Heart Rate Forecasting Under Scarce Ruck-March Data

Muntaqim Ahmed Raju, Ruizhe Ma, John W. Ramsay, David J. CornellarXiv 2026 · arXiv

Predicting how a soldier's heart will respond to the grueling demands of a ruck march could be a literal lifesaver, enabling real-time workload management and injury prevention. However, building these models is a classic "small data" nightmare: collecting high-quality physiological data during loaded marches is logistically difficult, leaving researchers with very few participants and sessions to work with.

A new paper from researchers at the University of Massachusetts Lowell and the U.S. Army DEVCOM Soldier Center tackles this head-on. Their core finding? To predict the individual, you must first learn from the crowd.

TL;DR

The study demonstrates that a "generalize then personalize" approach—pretraining a model on a broad population and then fine-tuning it for an individual—is the most effective way to forecast heart rate during intense ruck marching when personal data is scarce. This strategy significantly outperforms models trained only on population data or models built from scratch for a specific person.

What Problem Does This Paper Address?

Ruck marching—walking long distances with heavy packs—is a fundamental military task that places immense strain on the cardiovascular system. While wearable sensors can monitor heart rate (HR) in real-time, forecasting HR 30 to 60 seconds into the future provides a crucial "early warning" for fatigue or overexertion.

The challenge is twofold:

  1. Physiological Diversity: Every person responds differently to load based on their fitness, recovery state, and even the heat. A "one-size-fits-all" model often fails to capture these nuances.
  2. Data Scarcity: Unlike consumer fitness apps with millions of users, military-specific ruck-march datasets are tiny. Researchers often have to build models using data from fewer than a dozen individuals.

The authors set out to find the best training regime for this "scarce-data" environment: Should we rely on a general population model, try to build a custom model from scratch for each soldier, or find a middle ground?

Methods

The researchers benchmarked three distinct training strategies across three different neural network architectures (LSTM, Attention-LSTM, and CNN-LSTM):

  • Population-Only Training (POT): The model is trained on a group of people and tested on a completely new individual. It learns general trends but doesn't adapt to the specific person.
  • Personalized-From-Scratch Training (PFST): The model is trained only on the limited data available for the target individual.
  • Population-Pretrained Fine-Tuning (PPFT): The "Generalize Then Personalize" approach. The model is first trained on the population to learn the basic relationship between movement and heart rate, then "fine-tuned" using the small amount of data available for the specific individual.
Fig. 1. Methodological overview of the wearable-based HR forecasting benchmark for loaded ruck marching. Wearable inputs were converted into supervised forecasting sequences and evaluated using the persistence baseline, LSTM, Attention-LSTM, and CNN-LSTM models under POT, PFST, and PPFT. The decomposition pathway was evaluated only as a secondary 60 s ablation.

They evaluated these models on their ability to predict the HR trajectory over 30-second and 60-second horizons, using inputs like velocity, acceleration, and "Player Load" (a cumulative measure of physical work).

Datasets

The study used a specialized dataset of 8 active-duty U.S. Army medics performing 13 ruck-march sessions. Each session involved a 6.4-km march at a steady pace while carrying a 15.9-kg (35-lb) pack. While the number of participants is small, the high-frequency (1 Hz) wearable data provided thousands of overlapping windows for training.

Key Insights and Ideas

  • Strategy Trumps Architecture: One of the most striking findings was that the choice of training strategy (PPFT vs. PFST) had a much larger impact on accuracy than the choice of neural network backbone. Whether using a CNN or an LSTM, the "pretrain then fine-tune" approach was consistently superior.
  • The Failure of "Personalized-From-Scratch": In a data-scarce environment, trying to learn an individual's physiological signature from scratch is a losing battle. These models performed the worst because they lacked the "common sense" about HR dynamics that population data provides.
  • Persistence is a Formidable Foe: For short-term (30s) forecasting, a simple "persistence" baseline—predicting that the future HR will be exactly the same as the current HR—is surprisingly hard to beat in terms of average error. However, neural models (especially PPFT) proved better at capturing large deviations and trends at the end of the forecast window.
  • Transferable Structures: The success of population pretraining suggests that despite individual differences, there is a fundamental, transferable "HR–workload structure" shared across people during ruck marching.

Key Results

The Population-Pretrained Fine-Tuning (PPFT) approach was the clear winner across the board.

  • Error Reduction: Compared to training from scratch (PFST), the PPFT approach reduced the Mean Absolute Error (MAE) by 2.22–4.30 bpm.
  • 60-Second Horizon: At the longer 60-second window, the Attention-LSTM with PPFT achieved the lowest overall error (MAE of 5.573 ± 1.939 bpm).
  • Endpoint Accuracy: PPFT consistently achieved lower "endpoint MAE" (error at the very end of the prediction window) than population-only models, showing that personalization is most valuable for longer-range forecasts where individual drift becomes more pronounced.
Aggregate forecasting error at the 60-s horizon.

Limitations and Future Work

The authors acknowledge that their study is a "focused benchmark" on a small, all-male cohort.

  • Diversity: Future work needs to validate these strategies on larger, more diverse groups (including female participants) and across different environmental conditions like extreme heat.
  • Data Volume: The study didn't systematically test how much personal data is needed for effective fine-tuning. Is 5 minutes enough, or do we need 50?
  • Real-time Deployment: While the models are efficient, testing them in a truly real-time, "on-the-edge" wearable device remains a future step.

By proving that we can "borrow" knowledge from the group to better understand the individual, this research provides a practical roadmap for deploying advanced physiological monitoring in the most challenging field conditions.

Summary updated

Abstract

Heart rate (HR) forecasting may provide advance information about near-future cardiovascular response, potentially supporting physiological monitoring and workload-management decisions during ruck marching. However, robust model development is constrained by the difficulty of collecting large-scale ruck-march datasets, creating a scarce-data problem in which models must generalize across participants while adapting to individuals from limited personal data. To address this problem, we benchmark population-only training (POT), personalized-from-scratch training (PFST), and population-pretrained fine-tuning (PPFT) for HR forecasting during loaded ruck marching using data from eight active-duty U.S. Army medics across thirteen sessions. LSTM, Attention-LSTM, and CNN-LSTM backbones were evaluated against a persistence baseline at 30 s and 60 s horizons. Across all neural architectures and horizons, PPFT yielded lower mean MAE, RMSE, and endpoint MAE than both POT and PFST. Participant-level paired analysis after averaging across three training seeds showed that, relative to PFST, PPFT reduced trajectory MAE by 2.22-4.30 bpm and endpoint MAE by 1.24-2.88 bpm. All six trajectory-MAE comparisons remained significant after Holm adjustment ($p_{adj}$ = 0.0469), while four of six endpoint-MAE comparisons remained significant after adjustment. At 30 s, persistence achieved the lowest MAE (3.555 $\pm$ 1.274 bpm), while CNN-LSTM PPFT achieved the lowest RMSE (4.669 $\pm$ 1.916 bpm); LSTM PPFT achieved the lowest endpoint MAE (5.093 $\pm$ 2.365 bpm). At 60 s, Attention-LSTM PPFT achieved the lowest MAE (5.573 $\pm$ 1.939 bpm) and RMSE (7.134 $\pm$ 2.260 bpm), while CNN-LSTM PPFT achieved the lowest endpoint MAE (7.553 $\pm$ 2.357 bpm). These results suggest that population pretraining followed by participant-specific fine-tuning may provide a practical strategy for scarce-data wearable HR forecasting.
BibTeX
@misc{raju2026generalize,
  author = {Muntaqim Ahmed Raju and Ruizhe Ma and John W. Ramsay and David J. Cornell},
  title = {{Generalize Then Personalize: Wearable-Based Heart Rate Forecasting Under Scarce Ruck-March Data}},
  year = {2026},
  eprint = {2611.14849},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.14849},
}
Figure