Cardio Load: Contextualizing Cardiac Stress

• Cardio Load is a measure quantifying cardiovascular stress and workload, defined variously in exercise, cognitive, wearable, and biomechanical studies.
• In exercise and wearable contexts, CL uses heart rate, HRV, and activity data to estimate energy expenditure and adaptive training targets with robust performance metrics.
• In cognitive and biomechanical applications, CL leverages ECG-derived metrics and finite-element models respectively, highlighting both subject-specific calibration challenges and modeling assumptions.

Searching arXiv for papers on “Cardio Load” and related usages to ground the article. Cardio Load (CL) is a polysemous term in contemporary research. In the cited literature it denotes, depending on context, an internal training-load or energy-expenditure measure during exercise, a cognitive-load estimate inferred from cardiac dynamics, a wearable-derived score of cardiovascular work accumulated across the day, or a mechanical loading state of the myocardium defined by pressure, stiffness, and deformation. Across these usages, the common structure is the quantification of stress on the cardiovascular system or its signals, but the target variable, sensing modality, temporal scale, and mathematical formalization differ substantially (Gang et al., 2019, Meethal et al., 2024, Phillips et al., 15 Aug 2025, Mu et al., 24 Jul 2025).

1. Terminological scope and core distinctions

The term “Cardio Load” is not standardized across all subfields. In exercise-monitoring work it is used for in-exercise energy expenditure or internal training load, operationalized by calories or heart-response-derived stress. In cognitive-workload studies it denotes cognitive load inferred from cardiac activity, usually via ECG-derived HRV. In consumer wearables it is a proprietary training-load score based on heart rate reserve and accumulated minute-level cardiovascular work. In cardiac biomechanics it refers to the mechanical loading state acting on the ventricle, especially end-diastolic pressure and the resulting wall stress and strain (Gang et al., 2019, Meethal et al., 2024, Phillips et al., 15 Aug 2025, Mu et al., 24 Jul 2025).

Usage of CL	Operational target	Primary signals
Exercise physiology	Energy expenditure / training load	HR, HRV, activity metadata
Cognitive monitoring	Rest / low / high cognitive load	ECG-derived HRV, sometimes fused with pupil or EDA
Consumer wearable analytics	Cardiovascular work across the day	HR, HR reserve, inertial sensors
Cardiac biomechanics	Mechanical loading of LV geometry	Pressure, material parameters, mesh geometry

This multiplicity matters because claims about CL are often only valid within one of these formulations. A classifier that predicts calorie bands from post-exercise HRV is not estimating the same quantity as a weekly Fitbit CL target, and neither is equivalent to end-diastolic ventricular load. A plausible implication is that comparative discussions of CL require explicit identification of the latent variable being measured: physiological training stress, mental workload, population-facing training dose, or cardiac tissue loading.

2. Cardio Load as internal training load and energy expenditure

In “Prediction of Physical Load Level by Machine Learning Analysis of Heart Activity after Exercises,” CL is effectively the in-exercise internal training load expressed as calories burned, denoted $C$, with labels obtained from a commercial Firstbeat Technologies device (Gang et al., 2019). The study defines three load categories:

• Low load: $0 \le C < 400$ kcal
• Medium load: $400 \le C < 1000$ kcal
• High load: $1000 \le C \le 4000$ kcal

The central problem is to predict these calorie ranges from heart activity measured after exercise rather than during exercise. The reference in-exercise formulation is

$C \sim A + D + T + AHR + MHR + I + V + P$

with $A$ the type of activity, $D$ distance, $T$ duration, $AHR$ average heart rate during exercise, $MHR$ maximum heart rate during exercise, $0 \le C < 400$0, $0 \le C < 400$1, and $0 \le C < 400$2 (Gang et al., 2019).

The more distinctive post-exercise formulation is

$0 \le C < 400$3

where all features are measured in a 1-minute window after training, using a recording taken 30–60 minutes after exercise. A reduced model retaining the most important heart features is

$0 \le C < 400$4

(Gang et al., 2019).

The HRV feature set is classical time-domain HRV plus simple post-exercise heart-rate statistics. For a sequence of NN intervals $0 \le C < 400$5, the paper uses

$0 \le C < 400$6

$0 \le C < 400$7

$0 \le C < 400$8

$0 \le C < 400$9

$400 \le C < 1000$0

It also uses an HRV index defined as “the integral of the density of the NN interval histogram divided by its height,” together with RAHR and RMHR from the same 1-minute recording (Gang et al., 2019).

Seven algorithms are evaluated: Linear Regression, Linear Discriminant Analysis, k-nearest neighbors, Decision Tree, Random Forest, Gaussian Naive Bayes, and Support Vector Machine. For post-exercise prediction, the strongest reported results are obtained by k-nearest neighbors with the short model including activity type, with $400 \le C < 1000$1, $400 \le C < 1000$2, mean cross-validation accuracy $400 \le C < 1000$3, precision $400 \le C < 1000$4, and recall $400 \le C < 1000$5. With all heart features, Random Forest achieves $400 \le C < 1000$6 and $400 \le C < 1000$7 (Gang et al., 2019).

A key result is that removing exercise type $400 \le C < 1000$8 sharply degrades performance. In the short model without $400 \le C < 1000$9, k-nearest neighbors drops to $1000 \le C \le 4000$0, $1000 \le C \le 4000$1, mean accuracy $1000 \le C \le 4000$2, precision $1000 \le C \le 4000$3, and recall $1000 \le C \le 4000$4. This reflects the paper’s statement that swimming, cycling, and running show “significantly different patterns of energy expenditure and reaction of cardiovascular system” (Gang et al., 2019).

The study is a single-athlete proof of concept based on more than 300 training sessions over more than one year. It explicitly notes that the results cannot yet be generalized across individuals and that the models predict categories rather than exact calories (Gang et al., 2019). This suggests that, in this line of work, CL is best understood as a personalized latent mapping from post-exercise autonomic state to recent training stress.

3. Statistical, machine-learning, and fitness-oriented extensions

A related but distinct formulation appears in “Parallel Statistical and Machine Learning Methods for Estimation of Physical Load,” which does not use the term Cardio Load directly but operationalizes instantaneous and accumulated physical load from heartbeat or heart-rate distributions (Stirenko et al., 2018). The paper emphasizes heart beat (HB), i.e. RR interval in milliseconds, over heart rate (HR), and computes mean, standard deviation, skewness, and kurtosis on either accumulated ensembles or sliding windows such as “one hundred neighboring HB/HR measurements.”

With

$1000 \le C \le 4000$5

the paper uses

$1000 \le C \le 4000$6

and plots trajectories in a Pearson diagram with coordinates

$1000 \le C \le 4000$7

Rest-state HB distributions lie near the normal reference, exercise drives the distribution away from normal toward beta-type regions, and recovery moves trajectories back toward normal and uniform references (Stirenko et al., 2018).

The paper defines two geometric metrics on this diagram: Metric1, the distance from the normal distribution point, and Metric2, the distance from the uniform distribution point. Metric1 is proposed as a marker of growing physical load during exercise and decreasing load during recovery, while the moving-window versions capture local transitions and peaks (Stirenko et al., 2018). A plausible implication is that this framework treats CL as a distributional displacement from a rest-like cardiac state rather than as a direct energy-expenditure estimate.

The machine-learning component of that work predicts activity type from dynamic and heart-related features using multiple linear regression, a shallow neural network, and a deep neural network with four hidden layers trained by resilient backpropagation. The models show that adding heart features such as maximal HR and average HR improves classification relative to dynamics alone (Stirenko et al., 2018).

Another related line of work concerns cardio-respiratory fitness rather than sessional load. “UDAMA: Unsupervised Domain Adaptation through Multi-discriminator Adversarial Training with Noisy Labels Improves Cardio-fitness Prediction” and its later version “Turning Silver into Gold” address VO$1000 \le C \le 4000$8max prediction from free-living wearable data, using large silver-standard datasets and smaller gold-standard cohorts (Wu et al., 2023, Wu et al., 2022). In these papers, VO$1000 \le C \le 4000$9max is the benchmark metric of cardio-respiratory fitness, and the proposed UDAMA framework combines pre-training on noisy labels with adversarial domain adaptation. On the BBVS gold-standard cohort, the later report gives $C \sim A + D + T + AHR + MHR + I + V + P$0, correlation $C \sim A + D + T + AHR + MHR + I + V + P$1, MSE $C \sim A + D + T + AHR + MHR + I + V + P$2, and MAE $C \sim A + D + T + AHR + MHR + I + V + P$3 for UDAMA (Wu et al., 2022). The connection to CL is indirect but important: a more accurate estimate of fitness capacity provides a more physiologically grounded basis for interpreting training load relative to individual reserve.

4. Cardio Load as cognitive load inferred from cardiac activity

In cognitive-workload research, CL denotes cognitive load inferred from cardiac signals rather than exercise stress. “CALM: Cognitive Assessment using Light-insensitive Model” operationalizes cognitive load through a three-class $C \sim A + D + T + AHR + MHR + I + V + P$4-back design: Rest, CL1 as 1-back, and CL2 as 2-back, each lasting 3 minutes (Meethal et al., 2024). The cardiac modality is HRV derived from ECG, recorded simultaneously with pupillometry.

Two ECG sources are used: Biopac MP35, sampled at 1000 Hz, and Polar H10, sampled at 120 Hz. HRV features are computed on 60 s sliding windows with 10 s overlap, using neurokit2 for R-peak detection and RR-interval construction,

$C \sim A + D + T + AHR + MHR + I + V + P$5

The feature set includes mean RR interval, median RR interval, SDNN, RMSSD, pNN50, respiration rate, HF power, and LF/HF ratio (Meethal et al., 2024).

The paper’s central contrast is between light-sensitive pupillometry and light-insensitive cardiac features. Under light and dark conditions, mean pupil diameter differs with $C \sim A + D + T + AHR + MHR + I + V + P$6, whereas RMSSD differs with $C \sim A + D + T + AHR + MHR + I + V + P$7, indicating no significant light effect on that HRV feature (Meethal et al., 2024). Random Forest results show that multimodal HRV+pupil models substantially outperform pupil-only models. With all light conditions pooled, Random Forest reaches 89.42% accuracy for Biopac+Pupil and 92.26% for Polar+Pupil, versus 77.05% for pupil only and 80.96% or 82.15% for HRV only, depending on device (Meethal et al., 2024). The paper reports that multimodal data improves the accuracy by more than 20 percentage points over assessment based on pupillometry alone in some settings (Meethal et al., 2024).

The same general concept appears in “Measuring Cognitive Workload Using Multimodal Sensors,” where Easy versus Hard task difficulty is classified using ECG, EDA, respiration, and SpO$C \sim A + D + T + AHR + MHR + I + V + P$8 features extracted in 10-second windows (Hirachan et al., 2022). ECG features include BPM, IBI, stdI, stdD, RMSSD, PNN20, PNN50, MAD, and breathing rate. ECG alone reaches up to 0.68 accuracy with Decision Trees, while a fusion of ECG and EDA presents good discriminating power with accuracy $C \sim A + D + T + AHR + MHR + I + V + P$9 (Hirachan et al., 2022). Feature selection identifies ECG_PNN50, ECG_PNN20, ECG_IBI, ECG_MAD, and ECG_stdI among the most informative variables (Hirachan et al., 2022).

More recent work shifts from handcrafted HRV features to representation learning from raw wearable ECG. “CogAdapt: Transferring Clinical ECG Foundation Models to Wearable Cognitive Load Assessment via Lead Adaptation” adapts a clinical 12-lead ECG foundation model to 3-lead wearable cognitive-load assessment using LeadBridge, a learnable 3-to-12 lead adapter, and ProFine, a progressive fine-tuning strategy (Mousavi et al., 21 May 2026). Under leave-one-subject-out cross-validation, CogAdapt Scenario C reaches macro-F1 $A$0 on CLARE and $A$1 on CL-Drive, outperforming baselines trained from scratch (Mousavi et al., 21 May 2026). This line of work treats CL as a subject-independent classification target defined by wearable ECG windows and dense self-reports rather than explicit HRV summary statistics.

Across these studies, cardiac CL estimation is consistently framed as an autonomic-nervous-system readout. HRV features such as RMSSD, pNN50, SDNN, HF, and LF/HF are used because they reflect sympathovagal balance under mental workload (Meethal et al., 2024, Hirachan et al., 2022). A recurring theme is that cardiac measures are comparatively robust to optical confounds, especially changes in ambient light, and thus serve as a stable basis for multimodal workload assessment.

5. Consumer wearable Cardio Load and adaptive targets

A consumer-facing formalization appears in “Adaptive Cardio Load Targets for Improving Fitness and Performance,” which states that Cardio Load, introduced by Google in 2024, is a measure of cardiovascular work, also called training load, resulting from all the user’s activities across the day (Phillips et al., 15 Aug 2025). Unlike the calorie-band formulation of exercise studies, this CL is a dimensionless accumulated score derived minute by minute from heart rate reserve.

Heart rate reserve is defined as

$A$2

and intensity is expressed as percentage heart rate reserve,

$A$3

The per-minute Banister load is

$A$4

with $A$5 for males and $A$6 for females (Phillips et al., 15 Aug 2025).

The CL formulation modifies this TRIMP-style approach in three ways. First, CL accrues only once $A$7. Second, a minute must show evidence of movement from inertial sensors; minutes with elevated HR but no movement are discarded. Third, load in the 30–40% HRR band is down-weighted relative to the Banister curve, whereas above 40% HRR the full exponential weighting applies (Phillips et al., 15 Aug 2025). Daily and weekly load are then summed: $A$8

This formulation explicitly includes both workouts and incidental activity. The white paper notes that users averaging only about 10 minutes of workouts per day may still accumulate around 180 CL per week, with about 75% of that from daily non-workout activity, while very active users with at least 2 hours of workouts per day still obtain at least 20% of weekly CL from incidental activity (Phillips et al., 15 Aug 2025). Median weekly CL is reported as approximately 214 for males and 184 for females (Phillips et al., 15 Aug 2025).

A major extension is the introduction of adaptive weekly targets. Chronic load is estimated in two ways: a rolling mean over the previous 28 days,

$A$9

and an exponentially weighted moving average,

$D$0

The weekly target is then

$D$1

The rationale is to avoid dropping targets too quickly after brief interruptions while allowing targets to rise when higher training has been sustained (Phillips et al., 15 Aug 2025).

The same paper distinguishes CL from Active Zone Minutes (AZMs). AZMs are aligned with public-health guidelines and count moderate and vigorous minutes in a bucketed fashion, whereas CL uses continuous $D$2 with exponential intensity weighting, includes light activity down to 30% HRR, and is positioned as performance-oriented rather than guideline-oriented (Phillips et al., 15 Aug 2025). A common misconception is therefore that CL is simply another duration metric; in this framework it is specifically an intensity-duration integral with personalized scaling by heart-rate reserve.

6. Mechanical cardiac load in biomechanical modeling

In cardiac biomechanics, CL denotes mechanical load rather than training or cognitive demand. “HeartUnloadNet: A Weakly-Supervised Cycle-Consistent Graph Network for Predicting Unloaded Cardiac Geometry from Diastolic States” defines the relevant load state in terms of luminal end-diastolic pressure $D$3, resulting wall stress and strain, and the loaded end-diastolic geometry of the left ventricle (Mu et al., 24 Jul 2025). The unloaded cardiac geometry $D$4 is defined as “the state of the heart devoid of luminal pressure … a theoretical low-stress and zero-strain reference” (Mu et al., 24 Jul 2025).

The finite-element formulation minimizes total potential energy,

$D$5

where $D$6 is LV cavity pressure and $D$7 is the strain-energy density (Mu et al., 24 Jul 2025). The myocardium is modeled with a Fung-type transversely isotropic hyperelastic law,

$D$8

with $D$9 a global passive stiffness factor and $T$0 a quadratic form in Green–Lagrange strain components expressed in fiber coordinates (Mu et al., 24 Jul 2025). In this setting, higher pressure increases deformation from $T$1 to $T$2, while higher stiffness reduces deformation.

HeartUnloadNet learns the inverse map from loaded end-diastolic mesh to unloaded mesh, conditioned on physiological parameters such as ED pressure, myocardial stiffness scale, and fiber helix orientation (Mu et al., 24 Jul 2025). The model uses graph attention over arbitrary tetrahedral meshes and a cycle-consistency strategy linking unloading and re-loading. The total loss is

$T$3

with the cycle term enforcing consistency between predicted unloaded geometry and reconstructed ED geometry under the same loading parameters (Mu et al., 24 Jul 2025).

The training data comprise 20,700 FE simulations across 60 anatomical shapes and parameter grids including $T$4 mmHg and $T$5 Pa (Mu et al., 24 Jul 2025). Reported performance is sub-millimeter, with average DSC $T$6, HD $T$7 cm, inference time about 0.02 seconds per case, and more than $T$8 times speed-up relative to traditional inverse FE solvers (Mu et al., 24 Jul 2025).

This usage of CL is conceptually orthogonal to the wearable and workload literatures. Here CL is the combination of cavity pressure and material response that determines ventricular deformation. Its importance lies in providing a reference state for accurate stress and strain analysis, inverse parameter estimation, and intervention planning. A plausible implication is that “load” in cardiac modeling is best regarded as a boundary-value problem in continuum mechanics rather than a scalar training or workload score.

7. Synthesis, limitations, and recurrent themes

Across these literatures, CL is unified less by a single formula than by a recurring idea: cardiovascular state can be summarized into a load variable that supports inference, monitoring, or control. In exercise studies, the target is energy expenditure or internal training stress inferred from HR and HRV, often with activity type as a crucial conditioning variable (Gang et al., 2019). In statistical monitoring, the target is a deviation of heartbeat distributions from a rest-like reference, yielding instantaneous and accumulated fatigue markers (Stirenko et al., 2018). In cognitive-load research, the target is mental workload inferred from HRV or ECG representations, frequently with multimodal fusion to improve robustness (Meethal et al., 2024, Hirachan et al., 2022, Mousavi et al., 21 May 2026). In consumer wearables, the target is a daily and weekly training-load score based on heart-rate reserve, intensity-duration accumulation, and adaptive target setting (Phillips et al., 15 Aug 2025). In biomechanics, the target is the mechanical loading state that maps an unloaded ventricular geometry to a pressurized end-diastolic configuration (Mu et al., 24 Jul 2025).

Several limitations recur. Personalization is central in exercise-load inference, where single-subject studies explicitly warn against premature generalization (Gang et al., 2019). Cognitive-load studies report small cohorts, controlled laboratory tasks, or challenges in cross-subject generalization, motivating multimodal sensing and foundation-model transfer (Meethal et al., 2024, Hirachan et al., 2022, Mousavi et al., 21 May 2026). Consumer CL depends on accurate HR, valid $T$9 and resting-HR estimates, and motion detection; it is not a medical tool (Phillips et al., 15 Aug 2025). Biomechanical CL estimates depend on constitutive assumptions, homogeneous stiffness, and currently limited anatomical scope (Mu et al., 24 Jul 2025).

A common misconception is that CL always refers to a single physiological quantity. The literature instead supports four distinct but related meanings: internal training load, cognitive load inferred from cardiac activity, wearable cardiovascular-work accumulation, and mechanical cardiac loading. The most defensible general definition is therefore context-bound: Cardio Load is a model-based measure of cardiovascular stress or its correlates, whose operational meaning depends on the task, modality, and theoretical framework in which it is defined.