A Study of Machine Learning Models for Personalized

Heart Rate Forecasting in Mountain Biking

Xiaoxing Qiu, Jules White and Douglas C. Schmidt

Department of Computer Science, Vanderbilt University, Nashville, TN 37212, U.S.A.

Keywords:

Heart Rate Forecasting, Personalization, Cycling Course, Deep Learning.

Abstract:

Heart rate forecasting in cycling is most effective when it is personalized and course-speciﬁc to account for the

inﬂuence of individual and terrain factors. This paper empirically assesses various personalized and course-

speciﬁc heart rate forecasting models based on four machine learning models, including random forest, feed

forward neural networks (FFNNs), recurrent neural networks (RNNs), and long short term memory (LSTM).

The mean square error (MSE) is selected as the metric for model comparison. The results of our experiments

show that despite the severely overﬁtted random forest models the LSTM models have the lowest MSE in the

heart rate forecasting on our test dataset.

1 INTRODUCTION

1.1 Background

Monitoring exercise intensity during cycling is impor-

tant. While insufﬁcient exercise may slow improve-

ment in an athlete’s performance, excess exertion can

result in over-training or even muscle damage. Over-

training in cycling can result in a debilitating syn-

drome that degrades the performance of cyclists for

several months and may ultimately result in failure

to meet competition goals (Gleeson, 2002). More-

over, over-trained athletes—especially those involved

in endurance sports like cycling—are susceptible to

infections and require signiﬁcantly longer time for

recovery than non-athletes (Gleeson, 2002; Nieman,

1994).

Several metrics are available to estimate exer-

cise intensity, including heart rate, power output,

and

V O

max. According to Jeukendrup et al’s

study (Jeukendrup and Diemen, 1998), heart rate is

a reliable indicator of cycling exercise intensity since

it is largely independent of a speciﬁc course. How-

ever, heart rate can be affected by terrain factors, such

as grade of the hill, which varies on different riding

courses.

Predicting heart rate at different points in time on

a cycling course is hard since heart rate is affected by

course-speciﬁc features, as well as other personalized

physiological factors. Research has shown that a cy-

clist’s heart rate drifts upwards after exercising for 20

to 60 minutes despite unchanged work loads (Le et al.,

2009), which is a condition called “cardiac drift.”

Cardiac drift is associated with an increase of

core body temperature during exercise (Dawson et al.,

2005; Collinson et al., 2001), which may cause ath-

letes to lower their speed to maintain their heart rate.

On the other hand, higher speed can yield a higher

heart rate in a given condition (Le et al., 2009). More-

over, heart rate responses vary with a cyclist’s individ-

ual factors, such as gender and age (Le et al., 2009),

as well as their cadence on different parts of a given

course.

It is hard to predict heart rate on a speciﬁc point

in a course at a target speed. Training plans therefore

focus on setting target heart rate or power goals (e.g.,

power applied to the pedals by the rider) and riding as

fast as possible without exceeding those targets. Ide-

ally, training plans could be developed where heart

rate at a given speed and point in a course is predicted

and riders are given a set of target speeds to ride in-

stead. Achieving this ideal, however, requires build-

ing heart rate forecasting models that are personalized

and course-speciﬁc, which yields the following bene-

ﬁts:

• Achieve Speed Goals without Over-training. A

personalized performance forecasting model for

speciﬁc courses is vital for cyclists to establish

achievable speed goals at different courses in ad-

vance and maximize training effects without over-

training.

Qiu, X., White, J. and Schmidt, D.

A Study of Machine Learning Models for Personalized Hear t Rate Forecasting in Mountain Biking.

DOI: 10.5220/0010630600003059

In Proceedings of the 9th International Conference on Sport Sciences Research and Technology Support (icSPORTS 2021), pages 87-94

ISBN: 978-989-758-539-5; ISSN: 2184-3201

• Improve Course-speciﬁc Estimation Accuracy.

Riders can accurately estimate how much faster

they could ride in different sections of the course,

yet still remain within heart rate targets. For ex-

ample, a rider who has only ridden a course at

70-80% of max heart rate could estimate how fast

they would ride at a race intensity of up to 90%

of max heart rate. This estimate could give them

a benchmark of what they could currently achieve

without actually riding the course and potentially

over-exerting themselves.

A biker may need to climb up on a steep uphill

trail on uneven terrain with 6-8in rocks, which re-

quires signiﬁcantly slower speeds and greater inten-

sity than a similar climb on smooth ground. Similarly,

a high-speed downhill ride on smooth terrain will

place less stress on a rider’s core muscle groups com-

pared to the same downhill speed across tree roots.

The individual terrain features, turns, gradients, and

other aspects of a course signiﬁcantly impact a rider’s

speed and intensity, but current models are mainly

course-independent.

To maximize their efﬁcacy, personalized and

course-speciﬁc heart rate models should be easily

trainable from limited data—ideally a single test ride

of a course by an athlete. If an athlete rides a course

multiple times at multiple target heart rates, they are

less likely to need a predictive model since they al-

ready know the course well. Therefore, new courses

(i.e., where the athlete has limited knowledge) are

those where personalized and course-speciﬁc predic-

tive models are most valuable. Ideally, a cyclist

should be able to ride a course once and then predict

how different target speeds would impact heart rate at

different points on the course.

1.2 Research Question

This paper presents our research on comparing per-

sonalized multivariate models to forecast the heart

rate of a cyclist on a speciﬁc course using data from

a single ride. These models consider course-speciﬁc

factors at each part of the course (such as the grade

of road and the altitude), as well as current rider de-

tails (such as the cadence), and then forecast the heart

rate of the athlete based on them. We compare the re-

sults of heart rate forecast by the following machine

learning models from a single ride and report which

models perform best:

• The ﬁrst type of model experimented with is ran-

dom forest, which is a traditional machine learn-

ing model. Forecasting results show that random

forest models have severe overﬁtting and there-

fore cannot be utilized in heart rate forecasting

from a single ride.

• To mitigate the random forest overﬁtting prob-

lem, we implement feed forward neural networks

(FFNNs). FFNNs do not exhibit overﬁtting prob-

lems on the dataset, though their forecasting ac-

curacy is lower. In particular, FFNNs do not

consider historical information when forecasting

heart rate.

• To address the limitations with FFNNs, we also

implemented two other types of neural networks:

simple recurrent neural networks (RNNs) and

long short term memory (LSTM) networks.

We empirically compare all these models via ex-

periments in our testbed. The results of these exper-

iments indicate that the LSTM models have the low-

est mean square error among machine learning mod-

els despite severely overﬁtted random forest models

and produce predictions that closely match real-world

heart rate sequences.

1.3 Paper Organization

The remainder of this paper is organized as follows:

Section 2 describes prior work on heart rate forecast-

ing; Section 3 reviews the data processing methods

and all the models applied in this paper, including ba-

sic concepts of each model and their pros and cons;

Section 4 describes how we built random forest, feed

forward neural network (FFNN), recursive neural net-

work (RNN), and long short term memory (LSTM)

models to forecast the heart rate of athletes on a spe-

ciﬁc route and then compares the best models among

these four types with two course-independent models

(FitRec (Ni et al., 2019) and Minmin’s LSTM-based

model (Luo and Wu, 2020)); Section 5 presents con-

cluding remarks and outlines future work.

2 RELATED WORK

Researchers have built various models to predict the

performance of elite cyclists. Le et al. (Le et al., 2009)

proposed a mathematical model to evaluate athletes’

heart rate response under moderate exercise intensity

based on physical and physiological principles. Lu-

cia et al. (Luc

ıa et al., 2001) analyzed the preferred

cadence of elite cyclists and found that on ﬂat stages

they tend to adopt higher cadences (around 90 rpm)

while on mountain ascents cadences are around 70

rpm. However, their models focus on laboratory con-

ditions. Course-speciﬁc factors, such as the slope of

the road in real courses, are not consider, though these

icSPORTS 2021 - 9th International Conference on Sport Sciences Research and Technology Support

factors signiﬁcantly inﬂuence a cyclist’s heart rate re-

sponse.

Due to the advent of wearable devices, large

amounts of data can be collected and processed via

mobile devices, which offers an opportunity to build

personalized performance forecasting models. Xiao

and Ming et al. (Ming and Jun, 2008; Xiao et al.,

2010) used an FFNN to investigate the relationship

between heart rate and physical activity in daily life

with the help of a wearable physical activity recorder

that monitors the 3-D accelerations of the body.

RNNs exhibit sequential correlation and can

seamlessly model problems with multiple inputs.

These models are therefore widely used in natural

language processing and time series prediction (Cho

et al., 2014). In athletic performance forecasting,

RNN based models can take personalized factors

(such as blood pressure and running speed during ex-

ercise) to make a heart rate predictions. Ni et al. (Ni

et al., 2019) proposed an LSTM-based model to learn

a user’s heart rate proﬁle during exercise and offer

workout route recommendation and short term heart

rate prediction. Luo et al. (Luo and Wu, 2020) also

proposed an LSTM-based model to predict heat rate

based on heart rate signal, gender, age, accelerations

and mental state.

In summary, although there are other performance

forecasting models, most studies focus on the heart

rate proﬁle collected either during daily activities or

under laboratory conditions. There are few models

that are course-speciﬁc and personalized to forecast

a cyclist’s heart rate. However, such models can be

beneﬁcial to both cyclists and coaches.

For example, a cyclist needs a model to establish

various speed goals at different parts of a course be-

fore a competition or predict their heart rate based on

given speed goals. A coach can use such a model to

evaluate the heart rate of a given athlete on a given

course at a given speed to ensure the exercise inten-

sity and avoid over-training. Likewise, when given

speciﬁc heart rate goals, these models can be used to

predict how fast a cyclist can/should ride at different

parts of a course.

3 METHODOLOGY

To address the limited understanding of personal-

ized and course-speciﬁc heart rate forecasting from

a single ride, we evaluated the performance of prior

course/cyclist-independent models, course/cyclist-

speciﬁc traditional machine learning models, and

course/cyclist-speciﬁc neural networks on forecasting

rider heart rate on a single ride of a course. These

comparisons allowed us to investigate a number of

important research questions and collect important

lessons learned to guide future research, as discussed

in this section.

3.1 Key Research Questions

The key research questions we investigated in our

study include the following:

• Are current models that do not consider course-

speciﬁc features as good as models that consider

speciﬁc features, such as location?

• Which machine learning approaches perform best

for personalized course-speciﬁc heart rate predic-

tion from a single ride?

• For traditional machine learning models, which

features are most salient for learning?

3.2 Experimentation Approach

We began our study by surveying prior work on heart

rate forecasting. We then selected and applied both

traditional machine learning models and neural net-

works to a cycling dataset that we collected. Impor-

tant cyclist and course-speciﬁc factors must be con-

sidered for cycling performance forecasting. It is nat-

ural to consider multivariate models for performance

forecasting, e.g., random forest and neural networks

are popular machine learning algorithms because they

work for both regression and classiﬁcation and can

handle multiple inputs.

According to Leijnen et al. (Leijnen and van Veen,

2020), there are 13 major neural network architec-

tures used by researchers. Among all these neural net-

works, three of them are widely used in performance

forecasting, including feed forward neural networks

(FFNN) (Ming and Jun, 2008; Xiao et al., 2010),

basic recurrent neural networks (RNN) (Chowdhury

et al., 2019), and long short term memory (LSTM)

(Bian et al., 2019; Ni et al., 2019) models. We there-

fore selected these three neural networks and random

forest and built a personalized model to predict the

performance of a cyclist on a speciﬁc routes. LSTM

is a special type of RNN, so to distinguish these two

models we call basic RNN models ”simple RNN”

models in this paper.

Pruning irrelevant features to an athlete’s perfor-

mance can reduce model overﬁtting and improve fore-

casting accuracy. Traditional machine learning mod-

els such asrandom forest models are sensitive to data

variation, so small differences in the dataset can cause

a large variance in the prediction. Neural networks

have built-in mechanisms to mitigate overﬁtting and

A Study of Machine Learning Models for Personalized Heart Rate Forecasting in Mountain Biking

perform feature selection by assigning signiﬁcant fea-

tures larger weights. This process, however, con-

sumes a large amount of time and requires a large

amount of data. Removing less important factors can

accelerate the training process of traditional machine

learning models and neural networks and reduce the

amount of data needed. This paper therefore uses fea-

ture selection for all four types of models.

In statistics, a correlation coefﬁcient is used to

characterize how strong a relationship is between two

variables. Two common correlation coefﬁcients are

widely used: the Pearson correlation coefﬁcient and

the Spearman’s rank correlation coefﬁcient. Accord-

ing to Bishara et al. (Bishara and Hittner, 2017), cal-

culating Spearman’s correlation coefﬁcient for non-

normal data may be an optimal strategy when the

data size is larger than 20. Therefore, for heart rate

forecasting, Spearman’s rank correlation coefﬁcient

is more suitable for calculating the correlation coef-

ﬁcient, so we therefore select this approach to ﬁlter

the personalized and terrain factors.

3.3 Overview of Our Dataset

For heart rate forecasting, the dataset used in this pa-

per contains the grade of course, speed, heart rate, al-

titude, cadence, and distance at each second. We are

interested in understanding how course-speciﬁc fea-

tures impacted forecasting performance. We therefore

use mountain biking data from trails in the Nashville,

Tennessee, USA region.

Our dataset was collected on a Ripmo AF moun-

tain bike instrumented with a Garmin 830 biking

computer connected to an accelerometer-based speed

sensor mounted to the front hub of the bike. The

bike’s crank arms included a Quarq Sram XX1 Eagle

Dub power meter that used embedded strain gauges

to measure the power applied to the pedals within +/-

2%. The Quarq power meter directly measured the

rider’s pedaling cadence from accelerometers embed-

ded in the crank arms. The Garmin 830 included GPS

positioning and improved location tracking using a

fusion of wheel rotation, speed, and GPS ﬁx data. Fi-

nally, a Wahoo Tickr elctrode-based chest strap was

used to measure heart rate and communicate the data

to the Garmin 830.

For our study, we selected 8.71 miles of riding on

a 10-mile courses in the Nashville area. The rider

was a 40-year old male weighing approximately 210

pounds.

The dataset and all the source code we

Our current work focuses on a single rider to maxi-

mize understanding of how course-speciﬁc features could

be learned since individual riders have been more thor-

oughly studied in prior work.

used to evaluate the machine learning and neural net-

work models discussed in this paper is available from

github.com/EricXQiu/SportDataProcessing.

3.4 Feature Selection

The ﬁrst step in our dataset processing involved se-

lecting features for model training. Before selecting

these features their signiﬁcance must be determined.

Spearman’s correlation coefﬁcients (ρ) for each se-

quence are listed in Table 1. This table shows that the

Table 1: Spearman Correlation Coefﬁcient Between Factors

and Heart Rate.

Features

Spearman correlation

coefﬁcient

speed 0.1826

grade of course 0.2524

cadence 0.2389

distance 0.0466

altitude 0.2586

grade of course, cadence, speed, and altitude signiﬁ-

cantly inﬂuence heart rate more than the other factors.

We therefore selected these four factors as the features

for our heart rate forecasting model.

We split the dataset between a training set and test

set. A window average is carried out to metigate the

mesurement error of heart rate. By convention, we

used an 80% vs 20% train-test split ratio to split data

into the training set and test set. For the LSTM-based

model, the training set was the ﬁrst 80% of the data

and the test set was the remaining 20% rather than

a random selection to account for the order depen-

dence in heart rate data. To compare the results with

LSTM-based models, the same train-test split process

was also performed for the random forest model and

neural networks.

4 HEART RATE FORECASTING

MODEL COMPARISON

This section describes how we applied four machine

learning models (random forest, feed forward neu-

ral networks, simple recurrent neural networks, and

long short term memory) to build heart rate forecast-

ing models that predict an cyclist’s heart rate on a

given course. The results of applying these heart rate

forecasting models are then analyzed. We also com-

pare the mean squared error (MSE) of all four mod-

els to glean insights into which models perform best

and whether they have severe overﬁtting. Finally,

we compare our models with other models (i.e., Ni’s

icSPORTS 2021 - 9th International Conference on Sport Sciences Research and Technology Support

model (Ni et al., 2019) and Luo’s model (Luo and

Wu, 2020)) that are not course-speciﬁc, but which we

trained on our dataset described in Section 3.3.

4.1 Random Forest Heart Rate Model

We used the scikit-learn library (Pedregosa et al.,

2011) to construct the heart rate forecasting models

based on random forest. The criterion is MSE and

the number of decision tree regressors can be modi-

ﬁed. The number of decision trees regressors of four

random forest models are 5, 10, 15, and 20.

Each decision tree regressor uses all factors in

regression. Among them, model 1 with 5 decision

tree regressors shows the best performance. Its MSEs

are 0.455 bpm

and 125.15 bpm

on the training set

and test set respectively. The maximum depth of tree

depth is 31.

For all four models, the MSE on the test set is hun-

dreds of times larger than on the training set, which

shows that these random forest models have severe

overﬁtting. The depth of decision tree regressors is

approximately six times larger than the number of fac-

tors. This result indicates some factors are split mul-

tiple times, which increases the complexity of the de-

cision tree, thereby yielding severe overﬁtting of the

random forest model.

The heart rate forecasting result of model 1 is cho-

sen and shown in Figure 1. The predicted value fol-

lows the trend of the heart rate and shows the model

learned some patterns in the heart rate sequence due

to the course. However, the predicted heart rate shows

a large error around time 8,900s. Moreover, the pre-

dicted heart rate remains stable for the time range

from 8,600s to 8,700s, whereas the real heart rate

shows a sharp decreasing trend. The error in percent-

age between predicted heart rate and real heart rate

shows a similar trend as the forecasting result.

Figure 1: Heart Rate Forcasting of Random Forest Model.

To investigate the structure of random forest

model, the Gini importance

of each factor is calcu-

The Gini importance is the average decrease of vari-

ance, which indicates the probability of whether there is a

lated. A factor with large Gini importance means that

more nodes are split by this factor in the decision tree

and therefore this feature is considered signiﬁcant.

The average Gini importance of factors (in the or-

der of time, grade, speed, cadence, altitude) of model

1 is 0.294, 0.0376, 0.531, 0.028, 0.108 respectively.

It shows that speed, time, and altitude are three major

factors used to split nodes in the ﬁrst several layers

of decision tree. Likewise, the table also shows the

inﬂuence of cadence and grade are less important in

decision tree construction.

We hypothesize that the source of the error may

occur for the following reasons:

• The dataset only indirectly measures course fea-

tures via speed and position, so it does not ef-

fectively learn course-speciﬁc inﬂuences on heart

rate.

• There may be some internal rider conditions, such

as the mental activities of the cyclist over time.

For example, a cyclist may be anxious due to

the terrain difﬁculty, thereby increasing their heart

rate. All these features can inﬂuence heart rate,

though they are hard to measure and express quan-

titatively.

• The random forest models may overﬁt on the

training set and provide erroneous predictions.

4.2 The FFNN Heart Rate Model

Due to the overﬁtting problems of the random forest

models, we built feed forward neural network (FFNN)

models with different hyper-parameters. Two of them

have two dense layers and the other two have three

dense layers.

An important phase of training a neural network

involves ﬁnding a proper learning rate. To search

the learning rate systematically, we employed Smith’s

method (Smith, 2017) and utilized learning rate ﬁnder

implemented by Pavel Surmenok et. al (Surmenok

and Mackenzie, 2017). By running at each learning

rate for 10 epochs, the optimal start learning rate for

these four FFNN models is 3 × 10

−2

Figure 2 shows the predicted heart rate and the real

heart rate of the model applied to the same sample as

the random forest.

The sample shows that the FFNN model can fol-

low the trend of the cyclist’s heart rate on the speciﬁc

course. The FFNN model shows large errors from

8,500s to 9,000s. To investigate the error source, The

MSEs of the training set and test set are calculated.

The MSE on the training set and test set are 262 bpm

and 364 bpm

split on this feature (Menze et al., 2009).

A Study of Machine Learning Models for Personalized Heart Rate Forecasting in Mountain Biking

Figure 2: Heart Rate Forecasting of FFNN Model.

Compared with the random forest model, the vari-

ances of the FFNN model on the training set and test

set are much closer, indicating less overﬁtting sever-

ity. We therefore expect the FFNN model to general-

ize better than the random forest model because they

do not incur such severe overﬁtting.

4.3 The Simple RNN Model

Unlike FFNNs, recurrent neural networks (RNNs)

can utilize heart rate data in the past to forecast the

heart rate at the current moment. The architecture of

a simple RNN model for heart rate forecasting con-

sists two parts: 1. one or two Simple RNN layers 2.

several dense layers for output.

Simple RNN Models with different hyper param-

eters were built and four groups of hyper-parameters

are experimented. The models with one layer of Sim-

ple RNN layer perform dramatically better than mod-

els with two Simple RNN layers. The optimal start

learning rates were identiﬁed via Smith’s method and

an exponential decay learning rate function was uti-

lized for learning rate searching. With Keras learning

rate ﬁnder (Surmenok and Mackenzie, 2017), the op-

timal start learning rate was selected as 2 × 10

−1

The result of heart rate forecasting model with the

highest accuracy is predicted in Figure 3. The MSEs

on the training set and test set are 275.11 and 248.83

respectively. Compared with FFNN models, simple

RNN Models exhibit fewer errors and the forecasting

results are closer to the real heart rate sequence. The

reason of limited improvement may be ascribed the

gradient vanishing and gradient exploding problem in

Simple RNN.

4.4 The LSTM Model

Simple RNN models can suffer from vanishing gra-

dient problems, which limits their application when

the sequence of input data items is very long. To ad-

dress this issue, we implemented LSTM models. Four

sets of hyper-parameters are experimented and mod-

els with one LSTM layer model performs dramati-

Figure 3: Heart Rate Forecasting of Simple RNN Models.

cally better than models with two LSTM layers. This

may also indicate that the simple one-dimensional

signals of speed, position, and cadence are insufﬁcient

to learn course-speciﬁc heart rate variation.

With the help of keras learning rate ﬁnder (Sur-

menok and Mackenzie, 2017), the optimal start learn-

ing rate for models with one LSTM layer (i.e., Models

1 and 3) is selected to be 1 × 10

−1

The models with two LSTM layers show large

MSE on the test set and they do not learned the

general pattern of the heart rate compared with one

LSTMN layer (Model 1 and Model 3). The results

for these two one-layer LSTM models are shown in

Figure 4.

(a) Model 1

(b) Model 3

Figure 4: Heart Rate Forecasting of LSTM Models.

The MSEs on the training set and test set of LSTM

Model 1 is 141.17 bpm

and 200.50 bpm

. The MSEs

on the training set and test set of LSTM Model 3 is

62.47 bpm

and 196.61 bpm

icSPORTS 2021 - 9th International Conference on Sport Sciences Research and Technology Support

4.5 Comparison with Course

Independent Models

Based on the results presented above, it appears that

two LSTM models with only one LSTM layer per-

form better than the other three types of model. We

therefore also compared these LSTM models with

Luo’s LSTM model (Luo and Wu, 2020) and Ni’s

LSTM-based model (Ni et al., 2019), which are

course-independent and rely only on personalized

factors and contextual factors. The input features to

these latter two models were cadence, speed, altitude,

and time.

Since Luo’s and Ni’s models are not course-

speciﬁc, the grade of the biking course is excluded

in the input factors. In particular, only the structure

of their models are utilized and some layers (such as

the encoding layers in Ni’s model) are removed since

heart rate forecasting is the main focus. The struc-

ture of Ni’s model consists an LSTM layer, a dense

layer followed by a dropout layer while Luo’s model

includes two LSTM layers and two dropout layers.

These two models were ﬁrst trained on the same

training set as our heart rate forecasting models with

course-speciﬁc factors excluded. They were then

tested on the same test set. The MSE of all four of

these models is shown in Table 2.

Table 2: MSE of LSTM models.

Models

MSE on

Training Set

MSE on

Test Set

Ni’s model 1822.23 407

Luo’s model 239.89 157.40

LSTM model 1 141.17 200.50

LSTM model 3 62.47 196.61

The comparison of forecasted heart rate is shown

in Figure 5. While Ni et al.’s model showed less error

Figure 5: Comparison of Different Heart Rate Forecasting

Models.

on both the training set and test set, the heart rate it

forecasted is close to constant, except for several drop

outs. Luo et al.’s model captures the heart rate pattern

at around 170 beats per minute, but cannot follow the

heart rate pattern overall.

5 CONCLUDING REMARKS

This paper presents an empirical analysis of personal-

ized and course-speciﬁc models to forecast heart rates

for cyclists. In particular, we explored the perfor-

mance and feasibility of learning a cyclist’s course-

speciﬁc heart rate model from a single ride on a

given course. We implemented models using long

short term memory (LSTM), recursive neural network

(RNN), random forest, feed forward neural network

(FFNN), and previously published architectures and

compared their performance learning a cyclist. We

also compared our model with Ni et al’s (Ni et al.,

2019) and Luo et al.’s (Luo and Wu, 2020) models.

The heart rate forecasting results in Section 4

showed that the accuracy of two LSTM models out-

performed the other two types of neural network mod-

els and did not have as severe overﬁtting as the ran-

dom forest models. In Section 4.5, the LSTM course-

speciﬁc models also performed better than the two

LSTM-based course-independent models.

For research question 1 in Section 3.1 both course-

independent models do not capture the heart rate trend

of the cyclist well from 8,000s to 10,000s, as dis-

cussed in Section 4.5. In particular, the LSTM model

proposed by Ni et al. (Ni et al., 2019) mainly outputs

a rectangular heart rate sequence probably due to the

lack of course-speciﬁc factors, such as grade. Ni et

al. (Ni et al., 2019) state that their model focuses pri-

marily on short-term predictions (typically in a win-

dow of 10 seconds), so a 2,000-second sequence may

be too long for their model. Luo’s model forecasted

a somewhat continuous heart rate around 170 beats

per minute with small dips around 9000s and 9250s.

In general, our results show that these two models are

not as accurate as our LSTM models since they do not

consider course-speciﬁc factors.

For research question 2 in Section 3.1, random

forest models exhibit lower MSE on the dataset.

However, their substantial difference in MSE on the

training set and test set indicates that they all incur

severe overﬁtting. Moreover, as the data set grows

larger, the number of erroneous readings will un-

avoidably increase, thereby increasing the probability

of erroneous prediction due to the accumulated inﬂu-

ence of error. As a result, conventional machine learn-

ing models have difﬁculty on course-speciﬁc heart

rate forecasting.

For research question 3 in Section 3.1 speed has

the highest Gini importance, which aligns with prior

A Study of Machine Learning Models for Personalized Heart Rate Forecasting in Mountain Biking

work on heart rates for cyclists (Le et al., 2009) and is

utilized as the splitting factor for the ﬁrst layer. The

three major factors are speed, time, and altitude. In

contrast, the inﬂuence of cadence and grade are con-

sidered less important in decision tree construction.

Our future work focuses on scaling up our valida-

tion on a larger body of cyclists to determine whether

these results hold true across a range of riders. We

are also exploring how imagery of the course can aid

in understanding complicated course features, such as

terrain roughness. For personalized factors, we are

evaluating the extent to which learned course-speciﬁc

models transfer to other riders of the same gender and

age, as well as bike types. We are also considering

dynamic personalized factors, such as breathing rate.

For course-speciﬁc factors, image data and videos

are being collected and analyzed via neural networks.

We are analyzing roughness and course conditions at

different parts of the course from these images and

applying them in our heart rate forecasting model.

REFERENCES

Bian, M., Peng, B., Wang, W., and Dong, J. (2019).

An accurate LSTM based video heart rate estimation

method. In Lecture Notes in Computer Science (in-

cluding subseries Lecture Notes in Artiﬁcial Intelli-

gence and Lecture Notes in Bioinformatics), volume

11859 LNCS, pages 409–417. Springer International

Publishing.

Bishara, A. J. and Hittner, J. B. (2017). Conﬁdence intervals

for correlations when data are not normal. Behavior

Research Methods, 49(1):294–309.

Cho, K., Van Merri

enboer, B., Gulcehre, C., Bahdanau, D.,

Bougares, F., Schwenk, H., and Bengio, Y. (2014).

Learning phrase representations using RNN encoder-

decoder for statistical machine translation. In EMNLP

2014 - 2014 Conference on Empirical Methods in Nat-

ural Language Processing, Proceedings of the Confer-

ence, pages 1724–1734.

Chowdhury, S. S., Hasan, M. S., and Sharmin, R. (2019).

Robust Heart Rate Estimation from PPG Signals with

Intense Motion Artifacts using Cascade of Adaptive

Filter and Recurrent Neural Network. In IEEE Re-

gion 10 Annual International Conference, Proceed-

ings/TENCON, volume 2019-October, pages 1952–

1957.

Collinson, P. O., Boa, F. G., and Gaze, D. C. (2001). Mea-

surement of Cardiac Troponins. Annals of Clinical

Biochemistry: International Journal of Laboratory

Medicine, 38(5):423–449.

Dawson, E. A., Shave, R., George, K., Whyte, G., Ball,

D., Gaze, D., and Collinson, P. (2005). Cardiac drift

during prolonged exercise with echocardiographic ev-

idence of reduced diastolic function of the heart. Euro-

pean Journal of Applied Physiology, 94(3):305–309.

Gleeson, M. (2002). Biochemical and immunological

markers of overtraining. Journal of Sports Science

and Medicine, 1(2):31–41.

Jeukendrup, A. and Diemen, A. V. (1998). Journal of Sports

Sciences Heart rate monitoring during training and

competition in cyclists Heart rate monitoring during

training and competition in cyclists. Journal of Sports

Sciences, 16(1):91–99.

Le, A., Jaitner, T., Tobias, F., and Litz, L. (2009). A Dy-

namic Heart Rate Prediction Model for Training Opti-

mization in Cycling (P83). The Engineering of Sport

7, pages 425–433.

Leijnen, S. and van Veen, F. (2020). The Neural Network

Zoo. Proceedings, 47(1):9.

Luc

ıa, A., Hoyos, J., and Chicharro, J. L. (2001). Preferred

pedalling cadence in professional cycling. Medicine

and Science in Sports and Exercise, 33(8):1361–1366.

Luo, M. and Wu, K. (2020). Heart rate prediction model

based on neural network. IOP Conference Series: Ma-

terials Science and Engineering, 715(1):12060.

Menze, B. H., Kelm, B. M., Masuch, R., Himmelreich, U.,

Bachert, P., Petrich, W., and Hamprecht, F. A. (2009).

A comparison of random forest and its Gini impor-

tance with standard chemometric methods for the fea-

ture selection and classiﬁcation of spectral data. BMC

Bioinformatics, 10(1):213.

Ming, Y. and Jun, J. (2008). Heart rate prediction based on

physical activity using feedforward neural network.

Proceedings - 2008 International Conference on Con-

vergence and Hybrid Information Technology, ICHIT

2008, pages 344–350.

Ni, J., Muhlstein, L., and McAuley, J. (2019). Modeling

heart rate and activity data for personalized ﬁtness rec-

ommendation. The Web Conference 2019 - Proceed-

ings of the World Wide Web Conference, WWW 2019,

2:1343–1353.

Nieman, D. C. (1994). Exercise, infection, and immunity.

International journal of sports medicine, 15(3):S131.

Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V.,

Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P.,

Weiss, R., and Dubourg, V. (2011). Scikit-learn: Ma-

chine learning in Python. the Journal of machine

Learning research, 12:2825–2830.

Smith, L. N. (2017). Cyclical learning rates for training

neural networks. In Proceedings - 2017 IEEE Win-

ter Conference on Applications of Computer Vision,

WACV 2017, pages 464–472.

Surmenok, P. and Mackenzie, J. (2017). keras lr ﬁnder.

GitHub repository.

Xiao, F., Chen, Y. M., Yuchi, M., Ding, M. Y., and Jo, J.

(2010). Heart rate prediction model based on physical

activities using evolutionary neural network. Proceed-

ings - 4th International Conference on Genetic and

Evolutionary Computing, ICGEC 2010, pages 198–

201.

icSPORTS 2021 - 9th International Conference on Sport Sciences Research and Technology Support