Estimating running performance

Insights from data acquired in unconstrained free living settings

Marco Altini
10 min readApr 11, 2018

Our latest paper, titled “Estimating running performance combining non-invasive physiological measurements and training patterns in free-living” was accepted for publication at the 40th International Engineering in Medicine and Biology Conference.

In this work we built models able to estimate running performance (10 km time) using 2 years of real world data from more than 2000 individuals, including morning physiological measurements obtained using HRV4Training, workouts acquired from Strava and TrainingPeaks, anthropometrics and training patterns.

In particular, we provided insights on the relationship between training and performance, including further evidence of the importance of training volume and a polarized training approach to improve performance.

Check out all our research at this link.

A little background

At HRV4Training, our long term goal is to push further our knowledge on complex relations between physiological data, lifestyle and human performance.

Our approach consists in providing users with a clinical grade tool, and doing research on a larger scale compared to what is normally possible in laboratory settings.

After the initial validation [1] of our camera based HRV measurement, we have been exploring the relation between physiology and performance, in particular between estimated VO2max and running times in distances between the 10 km and the marathon [4].

In our latest paper, we decided to look not only at the relationships between these variables but also at the predictive power of different features sets we have engineered in the context of estimating running performance.

Small team, big hearts.

Back to running performance

Our motivations for this work are at least twofold. First, by estimating running performance we could provide individuals with better race pacing strategies, which are often guessed based on limited data. Secondly, we could also tailor training plans to individual abilities, therefore reducing injury risk.

Different anthropometric, physiological, and training characteristics influence human performance in running. Low body fat has been associated with better times [2], similarly to low resting heart rate (HR) and higher heart rate variability (HRV) [3]. Other physiological parameters measured in the lab, for example lactate threshold and VO2 max, have also been linked to better running times.

In our recent work we have shown how VO2 max estimated from running workouts highly correlates with running performance in events between the 10 km and the marathon [4]. Training related variables, such as training volume (distance per week), as well as average training speed have been associated to improved running performance too.

Recently, interest has shifted to training patterns analyzed over weeks or months. For example, most elite athletes train in a so-called polarized regime, in which most workouts are carried out at low intensities, and a few at very high intensity, as opposed to moderate intensity training, more typical of recreational runners [5], [6]. Even in recreational runners, a shift to a polarized training regime resulted in performance improvements [7].

a) Heart rate data for an individual training under a non-polarized approach (left) and an individual training under a polarized approach (right). Moderate intensity was defined as HR within 5% of a user’s average. b) Resulting training intensities distributions highlighting how a more polarized approach involves less time spent at moderate intensities. Here we set an arbitrary threshold of 30% of workouts carried out at moderate intensity, to assign a user to the polarized or non-polarized category, only for visualization purposes.

What’s new in our approach?

Most literature published on estimating running performance is constrained by small sample size and a rather homogeneous sample (for example only men, or a narrow age range or performance range). Variables included in the model are acquired under laboratory conditions or supervised settings that are not practical (VO2 max, lactate threshold, biomechanics [2]). Finally, parameters are analyzed in isolation (for example the impact of polarized training on performance [5]) and running time estimates accuracy is suboptimal or has not been cross-validated.

In the past few years, we have witnessed fast technological developments and integrations between different platforms and services (e.g. public APIs), resulting in increased availability of multivariate data streams acquired from mobile applications and wearable sensors (e.g. GPS, accelerometer, physiological data). Such developments are providing scientists with data at a scale that is typically not manageable in regular laboratory studies, and therefore with the opportunity of providing additional insights on the relation between physiology, training and performance. While several mobile applications and wearable sensors have been released on the market, providing users with metrics reflecting behavior (e.g. steps taken, distance ran, etc.) limited work has been carried out to provide insights on the individual’s performance ability outside of laboratory settings.

In this work we propose the first longitudinal, large scale analysis of running performance with respect to a wide set of variables either self-acquired or acquired automatically and non-invasively in free-living, without laboratory tests or supervision.

Analyzed variables include anthropometrics, resting physiology, training physiology, training volume, training patterns and previous performance.

The data

Data was collected using the HRV4Training app during 2016 and 2017 . A total of 2113 users (1891 male, 222 female) met the inclusion criteria: training with a heart rate monitor, linking the app via third party APIs to collect workouts data, including at least one 10 km, and taking at least a month of morning physiological measurements. A user’s best 10 km time was automatically identified as the fastest 10 km workout over the 2 years, and used as reference for this analysis. For each identified best 10 km time, the previous 3 months of data were used to extract features related to a user training volume and patterns. Totally, 464809 morning physiological measurements and 296739 running workouts were acquired during the 2 years longitudinal study, for an average of 220 morning measurements and 140 workouts with heart rate data per person.

Defining predictors of running performance

We computed features representative of different aspects that may contribute to running performance. Then, we created the following sets to analyze their impact on estimation accuracy:

  1. Ant : anthropometrics data. A user’s body mass index (BMI), age and gender.

2. Rest : resting physiological data (HR and HRV). HR was computed as the mean HR during the daily morning measurement, while as HRV feature we used the square root of the mean squared RR intervals difference (rMSSD), a marker of parasympathetic activity [3], [9].

3. Vol : training volume and speed. Average workout distance and speed.

4. TrPhy : physiological data during training. We computed the speed to HR ratio, a feature that relies on the fact that a more fit (faster) runner would maintain a lower HR while running at a certain speed, with respect to less fit (slower) runners. This parameter is the main predictor behind V O2 max estimation models relying on sub-maximal tests or workouts data [4], [10].

5. Pol : training polarization. Training polarization refers to training at different intensities, typically avoiding moderate intensity training. We derived features from workouts summaries to analyze the impact of training polarization on estimated performance. Features were: percentage of workouts performed at speeds 5% above or below a user’s average workout speed and the percentage of workouts where HR was within 5% of a user’s average HR rate, a feature used to represent lack of polarization.

6. Performance : past running performance. Finally, the last feature set included all previous features plus the best 10 km time found in the 3 months preceding the fastest time, as previous running performance is highly predictive of future running performance.

Examples of features used to estimate 10 km running performance.

What’s the accuracy?

Results were best for feature set Performance , with a RMSE of 2.68 minutes. Estimation was least accurate when using only anthropometrics data (set Ant , RMSE = 6.27 minutes) and improved progressively when adding resting physiological data (HR and HRV, set Rest , RMSE = 6.07 minutes), training volume and speed (average kilometers per workout and average speed per workout set Vol , RMSE = 4.04 minutes), physiological data during training (speed to heart rate ratio, set TrPhys , RMSE = 3.96 minutes) and training patterns (percentage of workouts at low and high intensities, set Pol , RMSE = 3.64 minutes).

Relation between predicted and estimated running performance and Bland-Altman plot for more than 2000 users that were part of this analysis. These are cross-validated results, meaning that no data used for model building was used for model validation.

What did we learn?

While no causal link can be established, as users did not participate in an intervention, it is of interest to determine the impact of features representative of training patterns as derived from workouts, for example training polarization, a hot topic these days.

In our analysis, age, BMI, resting HR, speed to HR ratio and time spent at moderate HR intensity entered the model with a positive sign, meaning that a lower value for these predictors is associated with a faster 10 km. On the other hand, HRV (rMSSD), average distance and speed, percentage of workouts performed 5% faster or 5% slower than the average training entered the model with a negative coefficient.

Thus, according to our dataset and analysis, a more polarized training regime, with a higher percentage of workouts preformed either faster or slower than the average workout, as well as a lower percentage of workouts performed at moderate HR intensity, is associated with improved performance.

What’s in it for you?

We have released a new feature in HRV4Training, lactate threshold estimation, basically turning around the modeling detailed in this post and published in the paper.

In practical terms, the lactate or anaerobic threshold, is approximately the pace you should be able to hold for a distance between 10 and 15 km. This is the criteria used in HRV4Training, which should help you making sense of the app estimation.

Intuitively, knowing your lactate threshold can help you defining pacing strategies for racing events between the 5 km and the half marathon (or longer, but in that case, other factors such as training volume start to play a more important role), as well as determining training pace for intervals and tempo runs. Some useful insights in this context are provided by Greg McMillan in this article, that I’d recommend checking out.

Lactate threshold estimation in HRV4Training and actual 10 km race time. The app provides a range considering the average error of the model, which is about 14"/km.

Closing remarks

In this work, we used data acquired longitudinally, in free-living, to provide accurate estimates of running performance on a dataset of 2113 runners of all levels. We investigated the relation between anthropometrics data, resting physiology, training patterns and performance, showing that running performance can be estimated accurately.

The estimation models developed in this work do not require laboratory tests, and could be practically employed by the growing community of recreational runners to estimate performance and tailor training plans.

With the release of the lactate threshold estimation in HRV4Training, we aim at doing just that.

Results for different feature sets are consistent with previous results from smaller studies [3], [7], [5], [4], showing a positive correlation between higher estimates of VO2 max, higher HRV, lower HR, higher training volume, higher training speed, a more polarized training regime and running performance.

Our analysis focused on readily available parameters that can be easily acquired and processed in free-living. However, more variables could be integrated as metrics linked to biomechanics and running power are also becoming available in the consumers market.

Additionally, results could be backed up by an additional laboratory validation. Future work will aim at both including more parameters as well as looking at performance changes over time to determine the estimation model’s ability to track such changes.


HRV4Training Research

HRV4Training Blog

Get the app for Android & iPhone.


[1] D. J. Plews, B. Scott, M. Altini, M. Wood, A. E. Kilding, and P. B. Laursen, “Comparison of heart rate variability recording with smart phone photoplethysmographic, polar h7 chest strap and electrocardiogram methods,” International Journal of Sports Physiology and Performance, pp. 1–17, 2017.

[2] P. Bale, D. Bradbury, and E. Colley, “Anthropometric and training variables related to 10km running performance.” British journal of sports medicine, vol. 20, no. 4, pp. 170–173, 1986.

[3] M. Buchheit, A. Chivot, J. Parouty, D. Mercier, H. Al Haddad, P. Laursen, and S. Ahmaidi, “Monitoring endurance running performance using cardiac parasympathetic function,” European journal of applied physiology, vol. 108, no. 6, pp. 1153–1167, 2010.

[4] M. Altini, C. Van Hoof, and O. Amft, “Relation between estimated cardiorespiratory fitness and running performance in free-living: An analysis of hrv4training data,” in Biomedical & Health Informatics (BHI), 2017 IEEE EMBS International Conference on. IEEE, 2017, pp. 249–252.

[5] J. Esteve-lanao, A. F. San Juan, C. P. Earnest, C. Foster, and A. Lucia, “How do endurance runners actually train? relationship with competition performance,” Medicine & Science in Sports & Exercise, vol. 37, no. 3, pp. 496–504, 2005.

[6] P. B. Laursen, “Training for intense exercise performance: highintensity or high-volume training?” Scandinavian journal of medicine & science in sports, vol. 20, no. s2, pp. 1–10, 2010.

[7] I. Mu˜noz, S. Seiler, J. Bautista, J. Espa˜na, E. Larumbe, and J. Esteve-Lanao, “Does polarized training improve performance in recreational runners?” International journal of sports physiology and performance, vol. 9, no. 2, pp. 265–272, 2014.

[8] “Hrv4training,”, accessed: 2018–01–20.

[9] M. Altini and O. Amft, “Hrv4training: Large-scale longitudinal training load analysis in unconstrained free-living settings using a smartphone application,” in 2016 38th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Aug 2016, pp. 2610–2613.

[10] M. Altini, P. Casale, J. Penders, and O. Amft, “Cardiorespiratory fitness estimation in free-living using wearable sensors,” Artificial intelligence in medicine, vol. 68, pp. 37–46, 2016.

Marco holds a PhD cum laude in applied machine learning, a M.Sc. cum laude in computer science engineering, and a M.Sc. cum laude in human movement sciences and high-performance coaching.

He has published more than 50 papers and patents at the intersection between physiology, health, technology and human performance.

He is the co-founder of HRV4Training and loves running.

Twitter: @altini_marco



Marco Altini

Founder, Data Science @ouraring Lecturer @VUamsterdam. PhD in Machine Learning, 2x MSc: Sport Science, Computer Science Engineering. Runner