A novel data-driven model to define physical activity metrics and predict mortality in patients with chronic obstructive pulmonary disease

Abstract

Physical activity (PA) is a well-established prognostic marker in Chronic Obstructive Pulmonary Disease (COPD). Traditional PA metrics, such as step count, often overlook movement intensity, while energy expenditure (EE) relies on indirect calorimetry assumptions. To address these limitations, we propose a data-driven PA metric that integrates movement frequency and amplitude derived from raw accelerometer data. This retrospective analysis, based on a Dutch COPD database, also evaluates the predictive value of the new score for mortality in COPD patients compared to step count and energy expenditure. Movement data and step counts were collected using McRoberts triaxial accelerometers. Fourier analysis was applied to extract movement frequency and amplitude, which were then used to compute the physical activity (PA) score. Kolmogorov–Smirnov test was conducted to assess whether the distributions of step count and PA score differed, followed by Kruskal–Wallis test to assess day-to-day movement variability. Logistic regression was used to evaluate and compare the predictive performance of the novel PA score against step count and EE. A total of 404 COPD patients (51.5% female; median [IQR] age: 57 [46–66] years) were included. The proposed PA score and step count exhibited similar daily patterns but differed significantly in distribution. Mortality data were available for 165 participants. The PA score achieved 2.8% higher accuracy and 5.6% higher balanced accuracy than step count in mortality prediction, while EE demonstrated the lowest predictive performance. The proposed PA score demonstrates stronger predictive power for mortality in COPD patients, highlighting the importance of integrating movement characteristics beyond simple step count, and offering a more refined metric for PA evaluation in both clinical and research settings.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-08614-x.

Introduction

Chronic Obstructive Pulmonary Disease (COPD) is a heterogeneous lung condition characterized by chronic respiratory symptoms (dyspnea, cough, sputum production and/or exacerbations) due to abnormalities of the airways (bronchitis, bronchiolitis) and/or alveoli (emphysema) that cause persistent, often progressive, airflow obstruction¹. It is a leading cause of mortality worldwide, accounting for approximately 3.2 million deaths in 2019², with this number expected to rise due to population aging and continued exposure to risk factors¹.

Physical activity (PA) impairment is a hallmark of COPD and plays a crucial role in disease progression and prognosis³. PA encompasses both functional capacity (the ability to perform physical tasks) and actual physical activity levels (the amount of movement performed over time)⁴. While various tests assess functional capacity in COPD patients⁵, the increasing use of wearable technology has facilitated more detailed assessments of PA levels in real-world settings⁶. However, the precise prognostic significance of different PA measures for mortality prediction remains uncertain^7,8.

PA in COPD patients is typically assessed based on duration, intensity, and sedentary time, which includes prolonged periods of sitting or lying down. Studies indicate that reductions in incidental daily activities can occur early in the disease course^9,10, sometimes even before patients experience significant breathlessness^9–11. A sedentary lifestyle can lead to a downward spiral, where inactivity results in physical deconditioning, further exacerbating exertional breathlessness and prompting individuals to limit both the duration and intensity of their PA^12–14. Given this cycle, several studies have explored PA as a predictor of mortality and hospitalization in COPD patients¹⁵.

PA assessment in COPD typically considers three key dimensions: movement intensity, duration, and count¹⁶. Intensity is commonly measured using metabolic equivalents (METs) ⁹, while duration reflects time spent engaging in PA, and count generally represents the number of movements¹⁷. In clinical practice, movement count (e.g., step count) is widely used due to its ease of measurement with wearable sensors^18–21. Many COPD patients reduce high-intensity activities early in the disease course, making walking the predominant form of PA in their daily lives¹⁰.

While higher daily step counts are generally associated with a lower risk of respiratory failure in COPD patients¹⁸, step count alone fails to capture movement intensity, a critical aspect of PA assessment. For example, walking at 4 km/h and running at 7 km/h may both result in 5,000 steps, but the intensity and metabolic demand differ substantially. Such nuances cannot be captured by step count alone.

An alternative metric, energy expenditure (EE), integrates movement count, duration, and intensity^22,23. However, EE estimation often relies on proprietary algorithms in wearable devices, limiting transparency and reproducibility in research.

Analyzing raw accelerometer data offers an opportunity to develop more precise PA metrics by directly quantifying movement frequency and amplitude. This approach can improve PA assessment accuracy, better differentiate activity levels, and ultimately enhance mortality prediction in COPD patients. Accordingly, this study has two primary objectives: (1) to develop a novel, data-driven PA metric (referred to as the PA score) that incorporates movement frequency and intensity from raw accelerometer signals, providing a more comprehensive assessment than step count alone; and (2) to evaluate the predictive performance of the PA score in comparison to step count and EE for mortality risk in COPD patients.

Methods

Study design and participants

This study is a retrospective analysis utilizing patient data from a Dutch COPD database. In this cohort, biomarkers were measured at a single time point, while movement data were recorded over a one-week period from 2014 to 2016. Mortality outcomes were tracked longitudinally with a follow-up period of 6 years. Physical activity data, including three-dimensional movement accelerations and step counts, were collected using Dynaport accelerometers (McRoberts, Netherlands) (https://www.mcroberts.nl/). The study cohort comprised 413 COPD patients from the Netherlands.

Physical activity data collection

Participants were instructed to wear the accelerometer on the lower back for at least 22 h per day over an eight-day period, removing it only for activities such as showering or swimming. This placement was selected to minimize movement restrictions associated with lower limb placement and to reduce motion artifacts commonly observed with upper limb placement²⁴.

The accelerometer recorded three-dimensional movement accelerations (FIG. S1) and step counts at a sampling frequency of 100 Hz. Data were considered valid if wear time exceeded 22 h per day for at least four days. Days with wear time below 22 h were marked as ‘invalid’ by the system. Movement accelerations from the first and last days were excluded due to insufficient data points. Patients wore the accelerometer during their routine daily activities as part of a broader study on physical activity and health outcomes in COPD.

Validation dataset

As an external validation measure, step counts were used to assess the accuracy of the proposed PA score in the absence of direct observational data (e.g., video recordings). Ideally, the movement count derived from the proposed model should align with step count measurements.

Data processing and PA score computation

To quantify PA, three-dimensional movement accelerations were combined into a single metric, the Euclidean Norm Minus One (ENMO)^25–27. The ENMO-derived signals were segmented by seconds and days, then denoised (FIG. S2) and analyzed in the frequency domain using Fourier analysis^28–30. A cutoff frequency of 5 Hz was applied to distinguish human movement signals, with higher frequencies attributed to noise³¹. This approach enabled the identification of individual movements (steps) along with their amplitude (intensity) (FIG. S3). Further methodological details are provided in the online supplement. The proposed PA score (FIG. S4) was computed as follows:

Daily PA score for patient on a day :

1

Overall PA score for patient :

2

where represents the movement amplitude corresponding to step count estimate in second of patient , denotes the total number of seconds the accelerometer was worn on day , and represents the study duration, excluding the first and last days. A higher PA score indicates greater physical effort exerted by the patient. The step count estimates () of all COPD patients were validated by comparing the step count recorded by McRoberts’ accelerometers.

Statistical analysis

To assess the consistency of movement patterns across different days of the week, the Kolmogorov–Smirnov test was used to determine whether the distributions of PA scores and step counts were statistically similar. Additionally, the Kruskal–Wallis test was applied to compare these metrics across days. To evaluate the prognostic value of the PA score, correlations between COPD patient characteristics and PA measures were analyzed. Key predictive features were selected for logistic regression, which was employed to predict mortality using three PA metrics:

• I.Step count
• II.Energy expenditure (EE) estimated by McRoberts³²
• III.Proposed PA score

Model performance was assessed using the following classification metrics:

• True positive rate (TPR)

• True negative rate (TNR)

• Positive predictive value (PPV)

• Negative predictive value (NPV)

• Overall accuracy (ACC)

• Balanced accuracy (BA)

All analyses, including validating the step count estimates, were conducted using MATLAB (R2023b, MathWorks, MA, USA) on a system equipped with an Intel® Core™ i7-10510U CPU and 16 GB RAM.

Results

Participant characteristics

The study cohort consisted of 413 COPD patients from outpatient clinics in the Netherlands (Table 1). According to the McRoberts guidelines, patients were required to wear the accelerometer for at least 22 h per day for four days to ensure data reliability. However, only 404 patients met this criterion and were included in the PA score computation. No additional information was available regarding participants’ daily routines or activities.

Table 1.

Participant characteristics.

	N = 404
Sex, n (%)
Male	196 (48.5)
Female	208 (51.5)
Age, years (IQR)	57 (46–66)
BMI, kg/m2 (IQR)	26.3 (22.7–29.5)

IQR – Inter-quartile range, BMI – Body Mass Index.

Among the 404 eligible patients, 51.5% were female. The median age was 57 years (IQR: 46–66), and the median BMI was 26.3 kg/m² (IQR: 22.7–29.5). Data on employment, marital status, education level, income, or ambulatory rehabilitation settings were unavailable. A total of 2,101 daily PA score measurements were recorded. The median PA score was 500.1 (IQR: 273.4–870).

Model validation

A high level of agreement was observed between movement counts derived from raw ENMO accelerometer data and the reference step counts provided by McRoberts, with 98.8% (2,075 out of 2,101 estimates) differing by no more than three steps (Fig. 1a).

Fig. 1.

Model validation.

Although the two-sample Kolmogorov–Smirnov (K-S) test confirmed that normalized PA scores followed a different distribution than normalized step counts³³, the Kruskal–Wallis (K-W) test indicated consistent PA scores across weekdays (Fig. 1b and c). Specifically, PA scores exhibited a trend similar to that of step counts, with higher values observed from Tuesday to Thursday compared to Friday through Sunday. While both PA scores and step counts also appeared higher on Monday than on Friday and Saturday, these differences were not statistically significant.

Application of the PA score

The applicability of the PA score is investigated by comparing mortality predictions for COPD patients over a 6-year period. Due to computer storage damage, mortality data were available for only 165 of the 404 patients, yielding a mortality rate of 17.6%. Among these 165 patients, 164 had complete data for FEV1%-predicted, 6-min walk distance (6MWD), and estimated energy expenditure (EE). Kaplan–Meier survival analysis of the 164 patients (Fig. 2) revealed that 93.8% survival probability beyond four years (95% CI 90.1%–97.6%), and 82.3% survival probability beyond five years (95% CI 76.0%–89.1%), as presented in Table 2.

Fig. 2.

Kaplan–Meier curve with associated 95% confidence intervals for a subset of 164 out of 404 COPD patients who had complete information on mortality, age, sex, BMI, 6MWD, and FEV1%-predicted.

Table 2.

Survival probability from Kaplan–Meier analysis and the associated 95% confidence interval.

Time (Day)	Number at Risk	Number Failed	Survival Probability	95% Confidence Interval Lower	95% Confidence Interval Upper
236	164	1	0.994	0.982	1
497	163	1	0.988	0.971	1
581	162	1	0.982	0.961	1
1025	161	1	0.976	0.952	1
1053	160	1	0.97	0.944	0.996
1151	159	1	0.963	0.935	0.993
1359	155	1	0.957	0.927	0.989
1381	153	1	0.951	0.918	0.985
1440	145	1	0.944	0.91	0.98
1464	140	1	0.938	0.901	0.976
1493	138	1	0.931	0.892	0.971
1496	136	1	0.924	0.883	0.966
1598	123	2	0.909	0.864	0.956
1640	121	1	0.901	0.855	0.95
1683	112	1	0.893	0.845	0.945
1685	111	1	0.885	0.835	0.939
1713	105	1	0.877	0.825	0.932
1738	102	1	0.868	0.814	0.926
1762	98	1	0.859	0.803	0.919
1773	96	1	0.851	0.793	0.913
1780	95	1	0.842	0.782	0.906
1783	94	1	0.833	0.771	0.899
1839	85	1	0.823	0.76	0.891
1856	82	1	0.813	0.748	0.883
1879	66	2	0.788	0.718	0.866
1885	62	1	0.775	0.702	0.856
1906	51	1	0.76	0.684	0.845

Number At Risk – The total number of survivors at the beginning of each period. The number at risk at the beginning of the first period is all individuals in the lifetime study. At the beginning of each remaining period, the number at risk is reduced by the number of failures plus individuals censored at the end of the previous period. Censoring means the total survival time for that subject cannot be accurately determined. In the context of this study, censored individuals were known to have survived until the end of the observation period, but no information was available regarding their status beyond that point.

Number Failed – The number of individuals who experienced the event of interest (typically death in this study) during a specific time interval.

Tables 3 summarizes the characteristics of the 164 COPD patients with complete data on FEV1%-predicted, 6MWD, and EE. The mortality rate in this subgroup is 17.6%. Among these patients, 44.5% are female, with a median age of 63.5 years (IQR: 58–70). The median BMI is 25.4 kg/m² (IQR: 21.8–28.9), and the median FEV1%-predicted is 55.4 (IQR: 45.1–68.9). The median 6MWD is 440.5 m (IQR: 371–505). Additionally, among the 164 patients, only 119 had measurements for the modified Medical Research Council (mMRC) dyspnea scale and 137 for GOLD stage classification. The median mMRC score is 2 (IQR: 1–3), and the median GOLD stage is 3 (IQR: 2–4). Notably, 111 patients had a GOLD stage of 2 or higher, indicating that the majority of the 164 patients exhibited at least moderate airflow limitation.

Table 3.

Characteristics of COPD patients for the 6-year mortality prediction.

	Total N = 164	Survivors N = 135	Non-Survivors N = 29
Sex, n (%)
Male	91 (55.5)	73 (54.1)	18 (62.1%)
Female	73 (44.5)	62 (45.9)	11 (37.9%)
Age, years (IQR)	63.5 (58–70)	62 (57–68)	70 (62.8–73.3)
BMI, kg/m2 (IQR)	25.4 (21.8–28.9)	25.4 (22.1–29.2)	24.7 (20.2–28.1)
FEV1%-Predicted, % (IQR)	55.4 (45.1–68.9)	56.8 (46.9–71.7)	47.4 (39.6– 58.2)
6-Minute Walk Distance, m (IQR)	440.5 (371–505)	460 (389.8–513)	350 (292–410.3)
mMRC Dyspnea Scale for 119 patients (IQR)	2 (1–3)	2 (1–3) for 99 patients	2 (2–4) for 20 patients
GOLD Stage for 137 patients (IQR)	3 (2–4)	3 (2–4) for 114 patients	4 (2–4) for 23 patients
Step Count (IQR)	6643.3 (4338.2–8918.9)	7387.4 (4968.8–9316.2)	4241 (2536–5935.9)
Average Step Count	6870.1	7331.9	4720.8
PA Score (IQR)	106.9 (62.8–159.8)	117.1 (72.6–173.9)	61.2 (30.6–91.2)
Average PA Score	122.4	134.7	65.3
Energy Expenditure, kcal (IQR)	2306.9 (1981.8–2596.6)	2334.7 (2017.5–2636.6)	2194.8 (1727.2–2444.3)
Average Energy Expenditure, kcal	2310.2	2341.0	2166.6
Overall Mortality Rate	17.7%	0%	100%

Total (164 patients) = Survivors (135 patients) + Non-Survivors (29 patients).

Figure 3 presents correlation analyses among key features (predictors) for the 164 patients.

Fig. 3.

Correlation heatmap of features (predictors). Spearman’s rank correlation coefficient r: 0 ≤|r|< 0.2 – very weak, 0.2 ≤|r|< 0.4 – weak, 0.4 ≤|r|< 0.6 – moderate, 0.6 ≤|r|< 0.8 – strong, |r|≥ 0.8 – very strong.

Three strong correlations were identified:

• BMI and EE (higher BMI correlated with increased EE)

• PA score and 6MWD (higher PA scores were associated with 6MWD)

• Step count and PA score (higher step counts corresponded to higher PA scores)

Other correlations were moderate or weak.

To assess the predictive utility of the PA score, logistic regression was performed to predict 6-year mortality based on the following predictor groups:

1. Age + 6MWD

2. Age + Step Count

3. Age + PA Score

4. Age + EE (baseline model for comparison)

There are three primary reasons for selecting these predictors. First, age, 6MWD, step count, and PA score were the only four features with non-zero predictor importance scores (Fig. 4), calculated using the minimum redundancy maximum relevance (MRMR) algorithm³⁴. This method selects features that are mutually dissimilar yet highly relevant to the outcome—in this case, mortality—yielding an optimal set comprising age, 6MWD, step count, and PA score. Second, age demonstrated a weak correlation with 6MWD, step count, PA score, and EE, ensuring its independent contribution to the model. Finally, PA score exhibited a strong correlation with both 6MWD and step count, reinforcing its relevance as a key predictor.

Fig. 4.

Predictor importance scores of the mortality prediction for COPD patients.

This study employed the standard 70–30 train-test split³⁵, assigning 115 patients to the training set and 49 to the test set. Key findings are summarized in Table 4. The Age + PA Score model demonstrated equivalent classification performance to the Age + 6MWD model, while outperforming Age + Step Count in several key metrics: sensitivity (TPR) by 11.1%, precision (PPV) by 16.7%, and BA by 5.6% in the test set. Additionally, the PA score achieved nearly 10% higher BA than the Age + EE model. Notably, the Age + EE model failed to identify any positive mortality cases, resulting in lower overall accuracy and BA compared to the Age + PA Score model. These findings suggest that the PA score provides stronger predictive power than EE estimated by McRoberts, underscoring its potential as a more reliable indicator for mortality prediction in patients with COPD.

Table 4.

Mortality prediction for COPD patients over a 6-year period. A positive case represents a mortality in 6 years.

Excluding GOLD Stage N = 164, 6-Year Mortality Rate 17.7%		TPR (%)	TNR	PPV (%)	NPV (%)	ACC (%)	BA (%)
Age + 6MWD	Training	10.0	95.8	33.3	83.5	80.9	52.9
	Test	22.2	97.5	66.7	84.8	83.7	59.9
Age + Step Count	Training	5	97.9	33.3	83.0	81.7	51.5
	Test	11.1	97.5	50.0	83.0	81.6	54.3
Age + PA Score	Training	10.0	95.8	33.3	83.5	80.9	52.9
	Test	22.2	97.5	66.7	84.8	83.7	59.9
Age + EE (baseline)	Training	0	100	NA	82.6	82.6	50
	Test	0	100	NA	81.6	81.6	50

TPR – True Positive Rate, TNR – True Negative Rate, PPV – Positive Predictive Value, NPV – Negative Predictive Value, ACC – Accuracy, BA – Balanced Accuracy.

NA – Not applicable due to failure to predictive any positive case.

The mortality prediction suggests that relying solely on step count or EE may underestimate mortality risk in COPD patients. The proposed PA score provides a more nuanced assessment of physical activity patterns, allowing for better identification of high-risk individuals.

The logistic regression coefficients for Age + PA Score are −0.52 for age, 0.0841 for the PA score, and a constant term of 2.9127. Therefore, the key formula for logistic regression can be summarized as:

3

A patient is predicted to be deceased if , otherwise, survival is predicted. It means age contributes negatively to survival, a higher age is associated with a higher mortality risk. In contrast, the PA score contributes positively to survival, with higher PA scores linked to a lower mortality risk. Therefore, lower PA scores are associated with an increased risk of death.

Discussion

This study presents a novel data-driven model that defines a personalized physical activity (PA) score for patients with Chronic Obstructive Pulmonary Disease (COPD). While step counts are commonly used to estimate PA in COPD patients, they fail to capture movement intensity (Table S1). The PA score proposed in this study incorporates both movement intensity and frequency, providing a more comprehensive measure of PA levels.

Higher PA scores and step counts on Monday–Thursday compared to Friday–Sunday

Figures 1b and c show that both PA scores and step counts were higher from Monday to Thursday than from Friday to Sunday. Several factors may explain this pattern:

• Pulmonary rehabilitation programs, physical therapy, and medical check-ups are typically scheduled from Monday to Thursday in the Netherlands, providing structured physical activity.

• Reduced access to rehabilitation services on Fridays through Sundays may lead to decreased activity levels.

• Cumulative fatigue over the week may result in lower PA scores from Friday to Sunday.

PA score calculation and validation

The data-driven approach for computing the PA score is not device-specific, it can be applied to accelerometer data from different devices. The PA score is calculated by combining step frequency and intensity, offering a more detailed measure of PA than step count alone.

The proposed model was validated using data from 404 COPD patients, resulting in a total of 2,101 daily PA scores. The weekday-weekend activity pattern observed in the PA scores closely aligned with step count trends, further validating its reliability as an effective PA indicator for COPD patients.

PA score vs. 6-min walk distance

As shown in Table 4, the combination of age and the proposed PA score yielded the same predictive performance for mortality as the combination of age and 6-min walk distance (6MWD). However, due to the absence of a second dataset for validation, it remains uncertain whether the PA score is a stronger predictor of mortality than 6MWD, or vice versa. Given that 6MWD is an established and widely validated prognostic marker in COPD populations^36–38, the PA score can, at minimum, be considered comparable in predictive value within the current dataset. Furthermore, the PA score offers several practical advantages. Unlike the 6MWD, which requires standardized testing conditions, clinical oversight, and patient cooperation, the PA score is derived from passively collected accelerometer data that reflects real-world, habitual physical activity. This facilitates continuous and remote monitoring of functional status, supports early identification of clinical deterioration, and reduces logistical burdens on both patients and healthcare providers.

Clinical Implications

Our findings indicate that the PA score provides a better assessment of mortality risk in COPD patients compared to traditional metrics such as step count and energy expenditure (EE) estimated by McRoberts. This enhanced predictive capability suggests several clinical applications:

• Early identification of high-risk patients: The PA score could help detect patients at higher risk of mortality, allowing for early or timely interventions such as pulmonary rehabilitation or closer monitoring.

• Integration into wearable technology: Incorporating the PA score into wearable devices or remote monitoring systems could enable real-time risk assessment and proactive interventions.

• Personalized COPD management: The ability to track PA variations over time could support personalized interventions aimed at improving long-term outcomes.

Methodological Limitations

Despite its strengths, this study has several limitations. First, the model does not differentiate between sedentary behaviors such as standing, sitting, or lying down, despite sedentary behavior being an independent predictor of mortality in COPD patients³⁹. Our dataset lacks the necessary information or video recordings to analyze these behaviors. Second, the model does not estimate movement speed, as raw accelerometer data alone cannot distinguish between fast and slow movements. Third, the model does not account for resistance exercise or estimate EE, as no corresponding measurements were available in the dataset.

Future Directions

To enhance the utility of the PA score, future research should focus on the following areas:

1. Validation in larger, independent cohorts:

Additional studies are needed to confirm its generalizability across diverse COPD populations.

• (2) Expanding the model with additional inputs:

Incorporating heart rate, oxygen saturation, or sedentary behavior indicators could enhance PA assessment.

• (3) Predicting COPD exacerbations:

The PA score could potentially be used to forecast COPD exacerbations based on trends in daily PA levels, similar to its use in mortality prediction.

Conclusion

This study demonstrates that raw accelerometer data can be transformed into a novel PA score that captures both movement intensity and frequency, offering a more comprehensive alternative to step count for assessing PA in COPD patients. Future advancements in data integration and predictive modeling may unlock its full potential in improving COPD patient outcomes through real-time monitoring and personalized interventions.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Author contributions

Martijn Spruit, Alex van ’t Hul and Qichen Deng designed the study. Alex van ’t Hul, Hieronymus van Hees, and Remco Djamin were responsible for the data collection. Qichen Deng was responsible for data analysis. Qichen Deng drafted the manuscript. All authors critically reviewed and revised the manuscript. All authors read and approved the final manuscript.

Funding

This study did not receive any funding for its design, analysis, and interpretation of data, or for writing the manuscript.

Availability of data and materials

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

The study design was reviewed by the Central Committee on Research Involving Human Subjects of the Netherlands. The medical ethical committee approval was waived by the Central Committee on Research Involving Human Subjects of the Netherlands because the study is non-experimental. All methods mentioned in the manuscript were carried out in accordance with the Declaration of Helsinki. No experiments were conducted on humans or animals, and no human or animal tissue was used in the study. he Central Committee on Research Involving Human Subjects of the Netherlands decided that informed consent from participants was unnecessary. The study was conducted in accordance with European Union directive 2001/20/EC and the Declaration of Helsinki. The Research Ethics Committee of the Radboud University Medical Centre, and Bernhoven Hospital reviewed and approved the study and considered that the study protocol did not fall within the remit of the Medical Research Involving Human Subjects Act (WMO) (ref: 2017/3597).