Interpretable Framework for Predicting Preoperative Cardiorespiratory Fitness Using Wearable Data

Abstract

Objectives:

Predicting preoperative cardiorespiratory fitness (CRF) is crucial for assessing the risk of complications and outcomes in patients undergoing surgery. CRF is formally evaluated through submaximal exercise testing with cardiopulmonary exercise testing (CPET) or the 6-minute walk test (6MWT). However, formal CRF testing is impractical as a preoperative screening tool. Wrist-worn devices with actigraphy and heart rate monitoring have become increasingly capable of predicting physiological measurements. Our aim was to develop a clinically interpretable machine learning (ML) model using wearable-derived physiological data to predict CRF for older adults, and to access whether this model can accurately estimate the 6MWT distances for preoperative risk evaluation.

Methods:

We examined heart rate and activity data collected from Fitbit devices worn by older adults (N=65) who were scheduled to undergo major noncardiac surgery. Data collection took place over a 1-week period prior to surgery while participants engaged in their typical daily activities. Our primary aim was to leverage this wearable technology to forecast CRF among this group. We employed a machine-learning ensemble regression model to predict CRF, using 6MWT outcomes as an index. Further, we applied a Shapley Feature attribution approach to gain insights into how specific features derived from wearable data contribute to CRF prediction within the model, aiding in personalized fitness prediction.

Results:

Adults with higher CRF exhibited elevated levels of moderate-to-vigorous physical activity (MVPA), maximal activity energy expenditure (aEE_max), heart rate recovery (HRR), and non-linear heart rate variability (HRV). These measures increased concurrently with improvements in 6MWT outcomes. Our regression models, employing random forest and linear regression techniques, demonstrated strong predictive capabilities, with coefficient of determination values of 0.91 and 0.81, respectively, for estimating CRF. The Shapley Feature Attribution approach elucidated those greater levels of MVPA, aEE_max, HRR, and nonlinear dynamics of HRV serve as reliable indicators of enhanced CRF test performance.

Conclusion:

The integration of wearable data-driven activity and heart rate metrics forms the basis for utilizing wearables to provide preoperative cardiorespiratory fitness assessments, supporting surgical risk stratification, personalized prehabilitation, and improved patient outcomes.

1. Introduction

Surgery ranks among the most invasive interventions in healthcare, contributing significantly to morbidity, mortality, long-term disability, and reduced quality of life, especially in older adults [1]. In the United States, over 4 million major surgeries are performed annually on this demographic, with long-term outcomes often being poor. For noncardiac surgeries, the 90-day mortality rate approaches 6% [2, 3]. A key factor contributing to the heightened surgical risk among older patients is the gradual reduction in their functional capacity as they age. Cardiorespiratory fitness (CRF) improves physical and cognitive function and is associated with a lower risk of cardiovascular disease [4], diabetes, and cancer [5]; fewer postsurgical complications [6]; and improved health-related quality of life [1, 7]. Enhancing surgical risk assessment for older adults is imperative for optimizing treatment outcomes.

Preoperative physical fitness is a major determinant of surgical outcomes and postoperative recovery, particularly in older adults [8, 9]. The gold standard assessment of cardiopulmonary exercise testing (CPET) is the measurement of peak oxygen uptake (VO2peak) at the maximum workload [10]. Another indirect CRF assessment, the 6-minute walk test (6MWT) has been clinically validated to determine the effects of therapeutic interventions [11] and prognosis [12, 13]. Previous studies have explored metrics like non-exercise testing cardiorespiratory fitness (NET-F) and heart rate over steps for wearable-based cardiorespiratory fitness assessment [14, 15]. Although the 6MWT is relatively easy to perform, it still requires trained personnel, in-person healthcare services and patient engagement to be performed accurately [16]. Therefore, it is crucial to develop effective and scalable methods for remotely evaluating functional capacity to screen and stratify the risk of surgical procedures among preoperative older adults in their everyday environments.

In recent years, wearable fitness monitors have emerged as promising tools for assessing CRF in diverse populations. These devices offer abundant dynamic physiologic and mobility data generated through multisensory systems [17–19]. Widely available and affordable, they accurately monitor aerobic exercise intensity, heart rate (HR) profiles, and estimate functional capacity in chronic disease management [16]. Additionally, commercially available activity trackers show potential for daily fitness monitoring, offering an objective measure for functional classification [20]. The majority of studies that have examined activity trackers have utilized a range of metrics including average HR, heart rate variability (HRV), step count, heart rate recovery (HRR) to physical stressors, and moderate-to-vigorous physical activities (MVPA) to evaluate salient factors such as fitness [21, 22], functional outcomes [16], frailty [23] and genetic risk of cardiovascular disease [21, 22]. However, currently, there are no wearable explainable machine model for interpretable personalized CRF prediction in the home environment for older adults undergoing major surgery.

Machine learning algorithms are gaining widespread adoption in healthcare due to their ability to rival or exceed human proficiency across various domains, such as medical AI [24–27], disease diagnosis [28–32], protein structure prediction [33], and medical concept retrieval [34]. A key factor contributing to their success is their domain adaptability, which has been enhanced by the advancement of sophisticated models and enhanced hardware capabilities [35]. However, this increased adaptability also renders these models opaque and demands model interpretability for all stakeholders, including patients, medical providers, regulators, and even the medical AI developers [36, 37]. Explainable artificial intelligence (XAI) approaches utilize, most popularly, local feature attributions [38] and the Shapley value [39, 40], which have emerged to address the issue of ML explainability by enhancing clinical interpretability [28, 31, 41, 42].

This manuscript aims to fill this gap by systematically evaluating the performance of wearable fitness monitors as a means of predicting preoperative CRF in older adults undergoing noncardiac surgery. In addition to providing accurate fitness predictions, this clinically interpretable ML model is designed to offer clinical explanations of personalized fitness matrices for clinicians, researchers, and healthcare technologists involved in preoperative assessment for older adults, supporting risk stratification, remote monitoring, and personalized prehabilitation planning. This study pioneers an interpretable preoperative CRF prediction using ML models and wearable fitness trackers. The key contributions can be summarized as follows:

• We investigated wearable data-driven preoperative HRV, maximal activity energy expenditure, and heart rate recovery metrics and their association with preoperative cardiorespiratory fitness test outcomes at home during typical daily activities.

• We developed a machine-learning CRF framework utilizing heart rate and activity data from fitness trackers to estimate outcomes of the 6MWT.

• We employed the Shapley feature attribution approach to interpret the heart rate and activity matrices within our CRF prediction model, delivering personalized explanations that are not only accurate but also clinically interpretable.

2. Experiment and Methodology

Our Wearable CRF Prediction Framework (Figure 1) integrates wearable data nodes, feature engineering, machine learning (ML), and interpretability to create a robust system for estimating 6MWT performance at home. The wearable data node represents the collection of preoperative fitness tracker data (heart rate sensing and actigraphy) at the highest sampling rate available from the device application programming interface (API). Following pre-processing and data segmentation, feature engineering is performed to develop the full set of physiological features evaluated in the model. The ML framework involves using machine learning techniques to estimate 6MWT performance based on the extracted feature set. Interpretation focuses on explaining how the model arrived at its predictions, enhancing interpretability after the ML model predicts cardiorespiratory fitness. Personalized CRF interpretation tailors its explanations and insights to individual fitness matrices.

Figure 1.

Comprehensive system architecture for predicting 6MWT outcomes using wearable fitness trackers and explainable AI techniques. (a) Data Node: Collection of heart rate and activity data from fitness trackers worn by preoperative patients. (b) Feature Node: Extraction of physiological and activity features from the wearable Fitbit data. (c) Machine Learning Framework: Application of machine learning algorithms to predict cardiorespiratory fitness (CRF). (d) Model Interpretation and Personalization: Explanation of the CRF prediction model and its application to personalized fitness explanations.

2.1. Study Design

We conducted this prospective cohort study at a single tertiary care medical center (Weill Cornell Medicine, New York, NY) between December 2021 and January 2024. Older adults (≥60 years of age) scheduled to undergo elective non-cardiac, non-neurosurgery with general or regional anesthesia anticipated for ≥ 2 hours were eligible to participate. Exclusion criteria included: 1) using a wheelchair; 2) ventilator-dependence; 3) room air oxygen saturation (<86%) at rest; 4) cognitive impairment, defined as an Abbreviated Mental Test score ≤ 6 (AMTS); 5) orthopedic impairment that compromises exercise performance; and finally, 6) inability to access a smartphone or tablet device capable of syncing to the Fitbit device. All procedures followed the ethical guidelines of the 1975 Declaration of Helsinki and were approved by the local Ethical Committee Weill Cornell Medicine has approved this study (IRB number: 21–04023531). Written informed consent was obtained from all enrolled participants.

Eligible subjects were identified by screening the surgical schedule at least ten days in advance. Study personnel contacted potential participants by phone to introduce the study, gauge interest, and confirm eligibility. Participation was voluntary, and participants were provided a wrist-worn Fitbit Charge 3 device (Fitbit, San Francisco, CA, USA) to wear for seven days before and thirty days after their procedure. Interested patients were pre-screened using the AMTS to identify pre-existing cognitive impairment. An AMTS score greater than 6 qualified them for the study, and informed consent was obtained electronically via REDCap [43] electronic data capture tools hosted at Weill Cornell Medicine. Participants were sent a Fitbit device via overnight shipping and received email instructions for setup and linking their Fitbit account to the research platform, Fitabase (Small Steps Labs LLC, San Diego, CA, USA). The study team provided phone support for device setup if needed. Participants were instructed to wear the Fitbit device for at least seven days pre-surgery. Before the day of surgery (DOS), study personnel coordinated with the pre-operative nursing team to avoid surgical delays. Participants arrived thirty minutes early for additional study procedures and completed baseline questionnaires remotely via secured REDCap, covering demographics, the Stanford Health Assessment Questionnaire Disability Index (HAQ-DI) [44], the Duke Activity Scale Index (DASI) [45], and the Geriatric Depression Scale (GDS) [46]. On the day of surgery, pre-operative assessments included the 3-minute Confusion Assessment Method (3D-CAM) [47], Handgrip Strength test [48], Fried Frailty Assessment [49], and the Six-Minute Walk Test [12]. The 6MWT was conducted in an operating room hallway with baseline vitals and Borg Scale ratings for shortness of breath recorded before and after the test. Participants were instructed to walk as far as possible on a flat surface in six minutes, in accordance with American Thoracic Society guidelines [50].

2.2. Fitbit Data Acquisition

Fitabase (Small Steps Labs, LLC, San Diego, CA), a cloud-based wearables research platform, was utilized for all Fitbit data collection, including actigraphy and heart rate recording. HR data are sampled at variable intervals by the Fitbit device, with readings taken every 3–15 seconds. To ensure consistency, HR data were resampled to maintain a constant rate of 10 samples per minute. Additionally, metabolic equivalent (MET) and calorie (estimated energy expenditure) data were recorded every minute. We excluded all data points where the MET was less than 1.

The intensity of physical activity was quantified using the MET scale, where different ranges corresponded to varying levels of activity [22]. Specifically, activities were categorized as follows: 1 to 3 MET represented light activity; 3 to 4.5 MET signified moderate-low activity; 4.5 to 6 MET indicated moderate-high activity, and activities exceeding 6 MET were classified as vigorous physical activity.

2.3. Clinical Data Acquisition

Prior to surgery, subjects were asked to complete baseline questionnaires and assessments regarding their medical history, cognitive status, and activity level. Subjects completed the HAQ-DI and M-DASI surveys. The M-DASI is a simplified 5-question screening tool that has moderate-to-low sensitivity and specificity in prediction of anaerobic threshold and VO2max. Study personnel also reviewed available electronic health records to obtain participants’ medical histories, including relevant laboratory results and current medications, and data points from the anesthesia record. On the day of surgery, enrolled subjects completed the 6MWT, [50] while being continuously monitored with a pulse oximeter and wearable fitness monitor.

2.4. Data Pre-Processing

In our study, Fitbit data exhibits variable sampling sizes on the Fitabase server. Specifically, MET and Calories (estimated energy expenditure) data are recorded at 1-minute intervals, while HR data is captured at intervals typically ranging from 5 to 10 seconds. To ensure consistency in our analysis, we resampled all HR data to a uniform sampling rate of 5 seconds, after filtering out HR data corresponding to MET values below 0. In free-living environments, activity epochs were identified based on predefined MET ranges. For HRV analysis [51], we selected uninterrupted HR data series longer than 5 minutes within the same individual activity class, as defined by the MET range. Additionally, we analyzed heart rate recovery (HRR) during high-moderate and heavy activity states. This involved considering an accelerating phase of at least 3 minutes. HRR was computed during the decelerating phase, providing insights into cardiovascular response post-activity. Detailed raw data sampling rate information can be found in Table S1.

2.5. Feature Extraction

We extracted wearable data-driven preoperative physical activity intensity (%), maximal activity energy expenditure, HRV, and HRR matrices during typical daily activities, as illustrated in Figure 1(b). Table S2 provides a comprehensive list of the feature names used in our study, along with their definitions and relevant descriptions.

2.5.1. Physical Activity (%)

Raw 3-axis accelerometry data are pre-processed by the Fitbit device, and only aggregated activity intensity and classifications are available as MET via the Fitbit API. All actigraphy data were processed to determine the duration spent at various activity intensities. The activity ratio is calculated by dividing the time spent in each activity category (light, moderate-low, moderate-high, and vigorous) by the total recorded activity time in minutes, using MET data.

$$\mathit{A}\mathit{c}\mathit{t}\mathit{i}\mathit{v}\mathit{i}\mathit{t}\mathit{y}(\mathit{\%})=\frac{\mathit{Time}\mathit{spent}\mathit{in}\mathit{specific}\mathit{activity}\mathit{level}(\mathit{minutes})}{\mathit{Total}\mathit{activity}\mathit{time}(\mathit{minutes})}\times 100$$ (1)

2.5.2. Maximal Activity Energy Expenditure (aEE_max)

aEE_max is the highest metabolic rate during intense activity, corresponding to peak HR. It’s estimated using a linear regression approach [52] based on simultaneous Fitbit-measured HR and Calories, with maximal HR calculated as (200–age×0.67) for women and (216–age×0.93) [53] for men without any exercise protocol. The corresponding calorie expenditure at maximum heart rate is defined as aEE_max, as illustrated in Figure 2(a). This study also calculated features, like regression coefficient (slope_HR/Calories), and vertical y-intercept.

Figure 2.

Sample feature extraction from fitness tracker data on heart rate and activity. (a) Maximal Activity Energy Expenditure (aEE_max): Measures the highest energy expenditure during activities. (b) Frequency-Domain Analysis of Heart Rate Variability (HRV): Analyzes the frequency components of HRV for insights into autonomic nervous system activity. (c) Poincaré Plot Analysis: Provides SD1 and SD2, reflecting short-term and long-term heart rate variability. (d) Detrended Fluctuation Analysis (DFA): Presents results for dfa_alpha1 and dfa_alpha2, indicating the scaling behavior of heart rate dynamics. Abbreviations: BPM (Beats Per Minute), HR (Heart Rate), HRmax (Maximum Heart Rate).

2.5.3. Heart Rate Variability (HRV)

In accordance with international guidelines [51], we conducted a standard HRV analysis using the Fitbit-derived raw HR time series data. This involved extracting time, frequency, and nonlinear domain HRV features from the HR data using pyHRV, an open-source python library [54] (refer to Table 1). Subsequently, approximately 100 HRV features were obtained from the entire HR recording for each subject. To ensure consistency, the HR data were segmented into 5-minute epochs based on a constant MET-defined activity level, which were then considered for HRV analysis.

Table 1:

Patient Demographics.

Characteristics	Study Cohort (n)
Age (Median, IQR, Mean, SD)	69.9, (65.6 – 75.2), 70.3, 6.4
Gender (Male %, Female %)	48%, 52%
BMI (Median, IQR, Mean, SD)	27.1, (24.7 – 30.1), 27.6, 4.5
AMTS Score (Mean, SD)	8.8, 0.8
Dyspnea (no SOB %, SOB w/ Moderate-Low exertion %)	82%, 18%
Smoking (Current %, Past %, Never %)	1.5%, 47.5%, 51%
Current Alcohol consumption (Yes %, No%)	57%, 43%
Comorbidities
Atrial Fibrilation (n, %)	5 (8%)
Arthritis (n, %)	11 (17%)
Peripheral artery disease (n, %)	0 (0%)
Coronary artery disease (n, %)	9 (14%)
History of stroke/cerebrovascular accident (n, %)	3 (5%)
Chronic steroid use past 30 days (n, %)	7 (11%)
Chronic opiod use past 30 days (n, %)	3 (5%)
Bleeding causing disorder (n, %)	3 (5%)
Anxiety (n, %)	7 (11%)
Bipolar or other mood disorder (n, %)	9 (14%)
Has a Pacemaker (n, %)	0 (0%)
Currently taking Beta Blocker (n, %)	18 (28%)
Disability Score [measured by HAQ-DI]
No incapacity [0 – 0.5] (n, %)	52 (80%)
Disability [0.5 – 1.5] (n, %)	12 (18.5%)
Severe Disability [1.5 – 3.0] (n, %)	1 (1.5%)
Estimate of Functional Capacity [measured by DASI]
DASI Score - total points (Median, IQR, Mean, SD)	50.7, (34.7 – 58.2), 45.1, 14
METs (Median, IQR, Mean, SD)	9, (7.0 – 9.9), 8.3, 1.8
Depression [measured by GDS]
Yes (n, %)	5 (8%)
No (n, %)	60 (92%)
Risk of Mortality [measured by CCI]
Mild [0–2] (n, %)	19 (29.2%)
Moderate-Low [3–4] (n, %)	27 (41.5%)
Severe [≥5] (n, %)	19 (29.2%)
Frailty [measured by FFI]
Robust [0] (n, %)	20 (30.8%)
Intermediate or Pre-Frail [1–2] (n, %)	39 (60%)
Frail [≥3] (n, %)	6 (9.2%)
Preoperative Variables
ASA Score (Median, IQR, Mean, SD)	3, (2 – 3), 2.6, 0.5
Anesthesia Type (general %, general + regional %)	62%, 38%
Subjective Judgement of Functional Capacity
Good [>10 metabolic equivalents] (n, %)	15 (23.1%)
Moderate-Low [4–10 metabolic equivalents] (n, %)	46 (70.8%)
Poor [< 4 metabolic equivalents] (n, %)	4 (6.2%)
Surgical Procedure
Urology (n, %)	29 (44.6%)
General (n, %)	21 (32.3%)
Transplant (n, %)	5 (7.7%)
Bariatric (n, %)	4 (6.2%)
Plastic (n, %)	2 (3.1%)
Gynecologic Oncology (n, %)	2 (3.1%)
ENT (n, %)	1 (1.5%)
Oral/Maxillofacial Surgery (n, %)	1 (1.5%)

2.5.3.1. Time Domain HRV Features

Time-domain HRV features were computed directly from the HR time series by deriving RR(or NN) intervals (60/HR) [24]. Several features were extracted through time domain analysis, including the mean of the RR intervals, the standard deviation of the RR intervals (SDNN), the root mean square of successive RR interval differences (RMSSD), as well as the percentage of successive RR intervals differing by more than 20ms and 50ms (pNN20, pNN50) [24, 55].

2.5.3.2. Frequency Domain HRV Features

Initially, the RR interval (RRI) data were transformed into equidistantly sampled data using cubic spline interpolation. Subsequently, a power spectral density (PSD) was computed for the RRI data utilizing Welch’s periodogram-based fast Fourier transform (FFT), as shown in Figure 2(b). Absolute powers within the very-low-frequency (VLF, 0–0.04 Hz), low-frequency (LF, 0.04–0.15 Hz), and high-frequency (HF, 0.15–0.4 Hz) bands were determined [24, 51]. Additionally, relative powers of the LF and HF bands in normalized units, along with the LF/HF power ratio, were calculated.

2.5.3.3. Non-linear HRV Features

Five nonlinear measures were utilized to assess the nonlinear dynamics within heart rate signals[24]. Sample entropy (sampen) was employed as a complexity quantifier for HR signals [51]. SD1 and SD2, derived from Poincaré plot analysis, were employed as nonlinear input features for HRV evaluation, as shown in Figure 2(c) [51, 55]. Additionally, Detrended fluctuation analysis (DFA) characterizes the self-similar properties of short-term HR signals, delineated by alpha₁, α₁ and alpha₂, α₂ representing short-range and long-range correlations, respectively, as shown in Figure 2(d) [51].

2.5.4. Heart Rate Recovery (HRR)

With the onset of exercise, HR increases due to parasympathetic withdrawal and sympathetic activation. After the cessation of physical activity, HR decays toward its pre-exercise level due to parasympathetic reactivation and sympathetic withdrawal [56]. Normally, in healthy individuals, HR recovers exponentially with a fast decrease during the first minute after physical activity, followed by a slow gradual decay until reaching the baseline HR. The HRR [56] over phases of T₆₀, T_120, T₁₈₀, T₂₄₀, and T₃₀₀ were extracted and denoted by HRR₆₀, HRR₁₂₀, HRR₁₈₀, HRR₂₄₀, and HRR₃₀₀ respectively.

2.6. Statistical Analysis

Statistical analysis was performed using SPSS 26 software (IBM, Armonk, NY, USA). One-way analysis of variance (ANOVA) was performed to identify the significant differences (significance level of 0.001) of features.

2.7. Building and Evaluating Regression Models

Our objective was to develop a machine learning (ML) regression model capable of accurately predicting 6MWT performance. Our modeling decisions were guided by the coefficient of determination (r-squared value) obtained from regression predictions. We utilized Scikit-learn version 1.3.1 for conducting the ML analyses [57]. These experiments were executed using Python 3.11 installed on a Linux server equipped with 26-core Intel E5–2680v4 CPUs, 768 GB of memory, and 4 Quadro RTX 6000 GPUs, each with 24 GB of memory. Data visualization was performed using the Seaborn and Matplotlib libraries to examine dataset distributions and generate visual representations of the results.

2.7.1. Feature Selection and Dimensionality Reduction

A hybrid approach, combining filter methods (SelectKBest) and decision tree-based random forest regressor, was implemented to identify the most important features correlated to regression outcome. SelectKBest is a feature selection method used to select the top k features based on their statistical scores [57]. Meanwhile, the random forest regressor assessed feature importance by leveraging a forest of decision trees. Our hybrid feature selection process can be represented by the following equations:

$$FeatureImportance(X,y,k{)}_{\mathit{SelectKBest}}={\mathit{a}\mathit{r}\mathit{g}\mathit{m}\mathit{a}\mathit{x}}_{\mathit{features}}(\mathit{s}\mathit{c}\mathit{o}\mathit{r}\mathit{e}(X,feature))$$

Where, $X$ represents the feature matrix; $\mathit{y}$ be the target vector; $k$ denotes the number of features to select; score ($X$, feature) computes the score for each feature in $X$; ${argmax}_{\mathit{features}}$ selects the top k features with the highest scores.

$$FeatureImportance(X,y{)}_{RFR}=\frac{1}{N_{\mathit{trees}}}\sum _{i=1}^{N_{\mathit{trees}}}\sum _{{\mathit{nodes}\mathit{\in }\mathit{tree}}_{\mathit{i}}}\mathit{ImpurityDecrease}(\mathit{node})\times \mathit{Proportion}(\mathit{node})$$

Where, $X$ represents the feature matrix; $k$ denotes the number of features to select; score ($X$, feature) computes the score for each feature in $X$; argmax_features selects the top $k$ features with the highest scores.

$$\begin{gathered} \mathit{Feature}\mathit{Importance}(X,k{)}_{\mathit{Hybrid}} \\ =\mathit{Feature}\mathit{Importance}(X,k{)}_{\mathit{SelectKBest}}\times \alpha +\mathit{Feature}\mathit{Importance}(X,k{)}_{RFR} \\ \times (1-\alpha ) \end{gathered}$$

Where, $\alpha$ is the feature selection weight, $\alpha \mathit{꞊}0.5$ in this study.

2.7.2. 6MWT Performance Estimation using ML Regression Model

Random forest (RF) is an ensemble learning method, constructing multiple decision trees through different data subsets, and voting on the results of multiple decision trees to get the output of the random forest [58]. RF regression analysis was performed to estimate the CRF outcomes in 6MWT using the heart rate and activity features, as shown in Figure 1(c). Scikit-learn ver. 1.1.3 was used for the machine learning analysis [57]. The model was analyzed by observing the root mean square error (RMSE) and correlation of determination (R²) between actual and estimated 6MWT measures.

2.8. Feature Attributions through Shapley Value

Explainable Artificial Intelligence (XAI) is a novel advancement within the ML community aimed at offering a comprehensible rationale for the predictions generated by complex, ‘black box’ ML models through feature attribution techniques [28, 37, 59, 60]. A commonly used feature attribution technique employs Shapley values to interpret complex models, assessing the average change in the model’s output when a feature is added to all other potential feature combinations [42]. The Shapley value, rooted in coalitional game theory, aims to equitably allocate the total surplus or reward achieved by a coalition of players to individual players within that coalition [39]. The mathematical expression of the Shapley value, ${\varphi }_i\left(x\right)$ of feature $i$ in predicting the output:

$${\varphi }_i\left(x\right)=\sum _{S\subseteq N\backslash \left\{i\right\}} \frac{\left|S\right|!\left(\left|N\right|-\left|S\right|-1\right)!}{\left|N\right|!}\left(f\left(x_s\cup \left\{i\right\}\right)-f\left(x_s\right)\right)$$ (8)

where $N$ is the set of all features; $S$ is a subset of features excluding $i$; $x_S$ represents the instance $x$ with only features in subset $S;f\left(x_S\right)$ is the model’s prediction for instance $x_S;f\left(x_S\cup \left\{i\right\}\right)$ is the model’s prediction for instance $x$ with features in subset $S$ plus feature $i$.

SHAP [42], a feature attribution approach using shapely value, was employed for interpreting the contributions of heart rate features in the 6MWT prediction model, as shown in Figure 1(d). Interventional TreeSHAP [60] is ideal for explaining Shapley values in tree-based ML models like decision trees, random forests, and gradient-boosting models. TreeSHAP stands out for being non-trivial, bias-free, and free from variance issues. SHAP dependence scatter plot (Figure 4) shows the effect a single feature has on the CRF predictions made by the model. A SHAP summary plot (Figure 5) represents how different features contribute to the final CRF prediction, specifically for a model’s overall performance [61].

Figure 4.

SHAP dependence scatter plot illustrating the individual impact of wearable data-driven features on CRF predictions. The SHAP dependence scatter plots show how each feature contributes to the predictions, including: (a) “sd_ratio_Moderate-High”, (b) “sampen_Moderate-High”, (c) “hr_max_Moderate-High”, (d) “aEE_MAX(J/min)”, (e) “Is patient on Beta Blockers”; Yes: 1, No: 2, (f) “fft_rel_vlf_Moderate-Low”, (g) “HRR_240”, (h) “sdnn_Moderate-High”, (i) “slope_HR/Calories”, (j) “Vigorous_Activity_ratio(%)”. (k) “dfa_alpha1_Moderate-Low”, and (l) “Moderate-High_Activity_ratio(%)”. The x-axis represents the value of the respective feature, while the y-axis denotes the SHAP value. A positive SHAP value suggests that the model is inclined to predict an increase in the 6MWT outcomes, whereas a negative SHAP value indicates a prediction of no increase in the 6MWT outcomes.

Figure 5.

SHAP violin summary plot visualizes the top 16 influential features of the Random Forest Regression (RFR) model for predicting 6MWT (6-Minute Walk Test) performance. Each violin plot represents a feature, with individual data points colored by their corresponding feature values from the training set. The x-axis of each point denotes the feature’s contribution to the final prediction of 6MWT performance. Positive SHAP values indicate features that support better CRF outcomes in 6MWT, while negative SHAP values suggest factors that could be detrimental to CRF outcomes in 6MWT.

3. Results

3.1. Participant recruitment and demographics

The study recruitment period extended from December 2021 to January 2024. Seventy scheduled surgical patients were assessed for eligibility, consented, and enrolled. Sixty-five patients completed the study, and 5 were excluded for surgical cancellations or rescheduling (n=3), technology issues (n=1), or opting out of wearable use (n=1). Table 1 presents the participant demographics, comorbidities, and preoperative clinical characteristics.

The study population had a mean age of 70.3 years (±6.4), with 48% males (n = 31) and 52% females (n = 34). The mean BMI was 27.6 (±4.5). The AMTS pre-screening tool showed a mean score of 8.8 (±0.8). Chart reviews indicated the following comorbidities among the 65 subjects: 28% (n = 18) were on beta blockers, 17% (n = 11) had arthritis, 14% (n = 9) had coronary artery disease, 14% (n = 9) had a recorded diagnosis of a mood disorder, 11% (n = 7) had anxiety, 11% (n = 7) were using steroids, 8% (n = 5) had atrial fibrillation, 5% (n = 3) had a history of stroke, 5% (n = 3) had a history of a bleeding disorders, and 5% (n = 3) were taking an opioid. No subjects had peripheral artery disease or pacemakers. A baseline Qualtrics survey revealed 18% (n = 12) experienced dyspnea after moderate exertion. Regarding alcohol consumption, 57% (n = 37) reported current use, while 43% (n = 28) did not. Smoking status revealed that 1.5% (n = 1) were current smokers, 48% (n = 31) were former smokers, and 51% (n = 33) had never smoked. Approximately two-thirds of the participants (n=40) were classified as ASA III, and all others (n=25) were ASA II.

3.2. Multivariate Regression Models Enable Cardiorespiratory Fitness Prediction

We built an ML model to predict CRF outcomes through 6MWT measures, utilizing wearable fitness tracker data. ML models were trained on Fitbit physiological and activity features. Key features included aEE_max, HRV, HRR, and the physical activity ratio (%), derived from MET and HR waveforms measured during typical daily activities.

We compared the performance of several ML regression models: (1) Random Forest Regressor (RFR), (2) linear regressor (LR), (3) Kernel Ridge (KRR), and (4) LASSO regressor (LASSO-R). For each model class, we performed extensive hyperparameter tunning (within the training set). LR, KRR, and LASSO-R models perform similarly (coefficient of determination, R² = 0.81). As a complex multivariate decision tree model, the RFR model performed best, showing the highest regression performance (R² = 0.91) to estimate the 6MWT measure, as a CRF indicator shown in Fig. 3. We chose the Random Forest Regressor for our final model over linear and non-parametric models (LR, LASSO-R, and KRR). A multivariate model like RFR offers a richer representation of the complex relationships among preoperative cardiac and activity features.

Figure 3.

Graphs illustrating the performance curves of machine-learning regression models employed for estimating 6MWT (6-Minute Walk Test) performance, showcasing the relationship between predicted values and actual outcomes. (a) Performance curve of Random Forest model. (b) Performance curve of Linear Regression model.

3.3. Explanation of Activity (%) as a CRF Metric

We explored non-linear and interaction effects in our features in the CRF prediction ML model. SHAP analysis of our CRF prediction model revealed a positive association between higher levels of vigorous and moderate-to-high intensity activity and greater CRF outcomes in 6MWT distance (Figures 4(j), 4(l), and 5). This finding aligns with established knowledge that regular physical activity, particularly MVPA, can increase CRF, and is cardioprotective in the primary and secondary prevention of cardiorespiratory diseases [22].

3.4. Explanation of Maximal Activity Energy Expenditure as a CRF Metric

SHAP analysis of the CRF prediction model identified aEE_max as a key factor associated with higher CRF as measured by 6MWT distance (Figure 4(d) and 5). This finding aligns with the understanding that greater aerobic capacity is linked to improved CRF. Additionally, the analysis revealed a shallower slope in HR versus calorie expenditure plot for adults with higher 6MWT scores (Figure 4(i) and 5). This suggests that individuals with higher fitness can sustain physical activity at higher intensity levels with less strain on their cardiorespiratory system [62, 63].

3.5. Explanation of HRV autonomic features as a CRF Metric

Through our investigation, we investigated both linear and non-linear techniques for HRV analysis across various activity levels, aiming to comprehend the autonomic tones and interaction effects within our features in ML models tailored for predicting fitness levels. Our analyses, depicted in Figure 4, revealed several intricate interactions.

HRV analysis revealed interesting associations with cardiorespiratory fitness as measured by the 6-minute walk test performance (Figure 4(k), 4(a), 4(b), and 5). Adults with higher 6MWT scores exhibited elevated non-linear HRV components during moderate-low and moderate-high intensity activities. Specifically, these individuals had higher dfa_alpha1_Moderate-Low, sd_ratio_Moderate-High, and sampan_Moderate-High values. Prior research suggests that these HRV markers, characterized by higher values, are associated with greater adaptability and complexity of the autonomic nervous system, which is often linked to greater CRF and overall health [62].

Analysis of HRV in the frequency domain yielded further insights into the association with CRF as measured by 6MWT performance (Figures 4 and 5). Individuals with superior 6MWT performance exhibited higher levels of specific frequency-domain HRV components, namely fft_total_Moderate-High, fft_total_Vigorous, fft_rel_vlf_Moderate-Low, and fft_rel_vlf_Moderate-High. This aligns with existing research suggesting that higher fft_total and fft_rel_vlf values in the frequency domain are associated with greater physical fitness levels.

3.6. Explanation of HR recovery as a CRF Metric

Results of HR recovery revealed a positive association with CRF as measured by the 6MWT performance (Figure 4(g) and 5). Adults with higher 6MWT scores exhibited faster HRR_240, indicating better cardiovascular health and fitness. This aligns with established knowledge that slower HRR is associated with increased risk of cardiovascular diseases and mortality, whereas faster HRR reflects a more efficient cardiovascular system [23, 64]. Enhanced CRF improves heart function and oxygen delivery during exercise, leading to faster heart rate recovery by reducing the workload on the heart [22].

3.7. Personalized Fitness Explanations through Shapley Value

Beyond predicting fitness test results, machine learning models with feature attribution capabilities in Random Forest Regression and Linear Regression can offer tailored insights into an individual’s cardiorespiratory fitness. This personalized explanation details how specific features derived from wearable sensor data influence CRF predictions within the models.

Figure 6 illustrates a Shapley force plot for a patient who achieved a 410.26-meter distance in 6MWT. The plot reveals that Moderate-High_Activity_ratio(%), dfa_alpha1_Moderate-Low, and sampan_Moderate-High emerged as protective factors for this individual’s CRF. In simpler terms, higher participation in moderate-to-high intensity activities, coupled with specific dynamic characteristics of HRV during moderate-low activities and overall activity patterns, contribute positively to their CRF level.

Figure 6.

Personalized fitness feature explanations for an individual patient: (a) SHAP waterfall plot illustrating heart rate trends as explanations for a ‘typical individual’ in predicting 6MWT performance, (b) SHAP force plot depicting personalized 6MWT performance calculation based on fitness risk and protective factors.

Conversely, the plot identified Vigorous_Activity_ratio(%) and aEE_max as risk factors. This suggests that lower levels of moderate-to-vigorous activity and a reduced maximum energy expenditure during exercise might be hindering the patient’s CRF. Based on these findings, it’s recommended that these individuals should prioritize improving moderate-to-high intensity activity levels while aiming for a higher maximum energy expenditure during exercise. This personalized approach highlights areas for improvement, ultimately promoting a more targeted path towards enhancing their overall cardiorespiratory fitness.

4. Discussion

This study investigated the potential of at-home wearable heart rate and actigraphy monitoring device to predict cardiorespiratory fitness in older adults scheduled to undergo elective noncardiac surgery. Importantly, we demonstrated that the model’s output, derived from wearable recordings collected on participants conducting routine, everyday activities, can serve as a surrogate measure of CRF. This presents a valuable tool for tracking perioperative fitness changes and informing individualized clinical planning. Wearable CRF measurements have potential to profoundly impact perioperative care with applications including goal-directed prehabilitation, surgery and anesthesia planning, early discharge planning, and postoperative remote patient monitoring.

The 6MWT is a submaximal exercise test, often used in cardiopulmonary patients to evaluate the physical functional capacity and distinguish symptomatic patients, provide prognostic information before therapeutic interventions, and thus can play an integral role in the preoperative decision-making processes [65]. A decrease in exercise capacity can indicate the onset of symptoms as well as a worsening of overall cardiopulmonary function and it is therefore commonly regarded as an indication for intervention prior to scheduled surgery [66].

4.1. Physical activity features as CRF predictors

Moderate-vigorous activity levels exhibit a positive correlation with the outcomes of the 6MWT, suggesting that individuals achieving higher distances tend to engage in increased levels of physical exertion, likely indicative of superior cardiovascular health and overall physical fitness [16, 21–23, 67]. Our findings indicate that participants with higher ratios of MVPA demonstrated superior 6MWT performance, which is likely associated with higher fitness levels. This association underscores the importance of engaging in activities that elevate heart rate and exertion levels, as they contribute to improved cardiorespiratory endurance and overall physical fitness.

4.2. Role of aEE_max as a CRF predictor

Among consumer-grade wearable fitness trackers, Fitbit may offer the most accurate measurement of energy expenditure, thanks to its well-regarded algorithm for estimating resting metabolic rate [68]. Our findings imply that subjects with higher maximal activity energy expenditure achieve better outcomes on the 6MWT. Individuals with greater fitness levels typically possess higher maximal activity energy expenditure capacities [69]. This is because enhanced fitness levels are indicative of improved cardiovascular efficiency, muscular endurance, and overall physiological adaptability. As a result, individuals with superior fitness can sustain higher levels of physical activity for longer durations without experiencing excessive fatigue or metabolic strain [70].

4.3. HRV Matrices and their role as CRF predictors

Moreover, HRV-based autonomic response measurements suggest increased variability in heart rate dynamics and can be used to assess CRF [62, 63, 71]. Our findings show that the non-linear HRV features, DFA_alpha1 and sd_ratio, indicate increased adaptability of the autonomic nervous system to physiological demands during moderate physical activity and correlate positively with cardiorespiratory fitness outcomes [72]. Additionally, individuals with better cardiorespiratory health often have elevated sample entropy measures, suggesting better modulation of HRV in response to exercise [63]. Moreover, we observed an upward trend in frequency-domain features, specifically the FFT total components, associated with higher CRF outcomes in 6MWT. Research indicates that individuals with higher CRF levels tend to exhibit increased variability in heart rate dynamics, reflecting a more robust and adaptable autonomic response to physiological stressors [55, 63, 71].

4.4. HRR features as CRF predictors

Further, studies have consistently demonstrated that adults exhibiting superior CRF outcomes in the 6-MWT tend to experience a faster HRR following exercise. This phenomenon underscores the close relationship between cardiovascular health and the body’s ability to efficiently regulate heart rate post-exertion [23, 64]. Our primary findings revealed faster HRR after exercise in adults exhibiting superior CRF outcomes in 6MWT. Individuals with higher CRF levels often display enhanced autonomic nervous system function, allowing for a more rapid return of heart rate to baseline levels after physical activity. Such expedited HRR not only reflects the efficiency of the cardiovascular system but also serves as a valuable indicator of overall cardiovascular health and fitness [22, 71].

The potential influence of age, medication, and physical fitness on heart rate poses a limitation when categorizing activity based solely on heart rate data. To address this concern, the use of beta blockers was incorporated as a covariate within the models. Interestingly, our analysis revealed that beta blocker usage did not have a significant impact on the achievement of target 6MWT distances or absolute distances. This finding suggests that while these factors may influence heart rate, they do not substantially affect the ability to attain desired walk distances, providing reassurance regarding the robustness of our models.

4.5. Strength and application

A key strength of our study, in addition to its considerable sample size, lies in our approach to translating wearable data-derived metrics into easily interpretable measures such as energy expenditure, HRV, HRR, and physical activity. We achieved this by developing equations based on data collected from combined heart rate and acceleration sensors during week-long daily living activities, without the need for a specific exercise protocol. This method allowed us to capture a comprehensive picture of participants’ physiological responses in their natural environments, enhancing the relevance and applicability of our findings. This fully transparent framework is therefore a first step towards personalized CRF prediction from wearables. This interpretable personalized framework provides insight into the protective and risk factors influencing CRF, highlighting the importance for the individual to prioritize improvements in these areas to enhance their overall fitness level. This approach could be readily adopted by device manufacturers, enabling users to access real-time feedback on their daily activity levels in the context of health outcome predictions. Such implementation holds the potential to empower individuals with actionable insights into their physical well-being, particularly for prehabilitation of older adults anticipating an elective surgical procedure.

4.6. Limitations and future directions

This study focused on predicting CRF outcomes in 6MWT using wearable activity and heart rate data. However, several limitations should be noted. Firstly, we used a few Fitbit-derived metrics like METs and HR, which rely on undisclosed, non-validated proprietary algorithms not widely approved for clinical use. Furthermore, raw accelerometer and ECG/PPG/Spo2 data are not accessible from Fitbit devices, restricting the ability to independently verify or improve upon the derived features. It is also likely that Fitbit’s algorithms were developed and validated primarily on healthy individuals, which may limit the reliability and generalizability of these estimates in clinical or patient populations. In addition, HRV metrics were derived from raw HR time series rather than direct beat-to-beat intervals from ECG or PPG, potentially introducing inaccuracies due to sensor lag, motion artifacts, and lower sampling resolution. Irregular HR sampling intervals of commercial fitness, such as, Fitbit devices (3–15 seconds) can introduce variability and reduce the accuracy of derived cardiac metrics compared to continuous, high-frequency Beat-to-beat cardiac profiles. Additionally, certain potentially informative fitness metrics, such as heart rate over steps, which could provide deeper insights into heart rate dynamics, were not included in this study due to the physical limitations of our elderly surgical patient population. Moreover, this study has not explored the relationship between predicted CRF and future health outcomes.

Future studies should consider incorporating these metrics to capture a broader range of physiological responses and improve the model’s predictive accuracy in more active or less frail populations. Moreover, future longitudinal studies are needed that include more detailed data points to establish a stronger link between our model’s learned representations and potential health risks. Finally, although most clinical wearable studies involve small sample sizes [73], validating the CRF prediction framework on a larger and more diverse sample size of pre-surgical patients would enhance its generalizability and clinical applicability.

5. Conclusion

This study investigated the associations between various physiological markers derived from wearables and cardiorespiratory fitness assessed by the 6MWT. Our findings demonstrate that higher activity energy expenditure, faster heart rate recovery, and specific characteristics of heart rate variability are positively correlated with better CRF. Beyond prediction, we showcase a methodology for translating readily available wearable data into meaningful fitness metrics linked to CRF risk. This framework paves the way for personalized preventive measures using wearables in everyday life. By empowering individuals with personalized insights into their cardiovascular health derived from wearable sensor data, this approach holds promise for personalized risk stratification and prehabilitation of older adults scheduled of elective surgery.

Highlights.

• Wearable fitness trackers predict preoperative cardiorespiratory fitness (CRF).

• ML models yield strong CRF predictions from wearable vital sign and activity data.

• Shapley Feature Attribution reveals MVPA, HRR, and HRV as top CRF indicators.

• The model aids preoperative CRF assessment and surgical risk stratification.

Data availability statement

The deidentified patient data may be available from the corresponding author upon reasonable request.