Associations between Performance-based and Patient-reported Physical Functioning and Real-world Mobile Sensor Metrics in Older Cancer Survivors: A Pilot Study

Abstract

Introduction:

Older cancer survivors are at increased risk for impaired physical functioning, but current assessments of function are difficult to implement in busy oncology clinics. Mobile devices measuring continuous activity and mobility in daily life may be useful for estimating physical functioning. The goal of this pilot study was to examine the associations between consumer wearable device (a wrist-worn activity tracker) and smartphone sensor data and commonly used clinical measures of physical function in cancer survivors aged 65 and older.

Materials and Methods:

Older adults within five years of completing primary treatment for any cancer completed standardized questionnaires and performance-based tests to measure physical functioning. Continuous passive data from smartphones and consumer wearable devices were collected for four weeks and linked to patient-reported and performance-based physical functioning as well as patient-reported falls or near falls at the end of the four-week monitoring period. To examine associations between sensor variables and physical functioning, we conducted bivariate Pearson correlations as well as multivariable linear regression analyses. To examine associations between sensor variables and falls, we conducted exploratory receiver operating characteristic curve and multivariable logistic regression analyses.

Results:

We enrolled 40 participants (mean age 73 years old, range 65–83; 98% White; 50% female). In bivariate analyses, consumer wearable device features reflecting greater amount and speed and lower fragmentation of walking in daily life were significantly related to better patient-reported function (r=0.43–0.65) and performance-based physical function (r=0.56–0.72), while smartphone features reflecting more geographic mobility were related to better performance-based physical function (r=0.40–0.42) but not patient-reported function. After adjusting for age and comorbidities, only consumer wearable device features remained associated with performance-based physical functioning. In exploratory analyses, peak gait cadence was associated with fall risk even after covariate adjustment.

Discussion:

This study provides preliminary evidence that real-world data from consumer devices may be useful for estimating functional performance among older cancer survivors and potentially for remotely and longitudinally monitoring functioning in older patients during and after cancer treatment.

Introduction

By 2030, 70% of cancer survivors will be 65 years or older. Cancer and aging appear to have additive effects on physical function, with older adults with a history of cancer exhibiting steeper functional declines compared to older adults without cancer¹. Approximately half of older patients with cancer show impaired physical function², and a review of 25 studies of older cancer survivors concluded that physical function is a robust and independent predictor of all-cause mortality³. Impaired physical function can also interfere with ability to work, care for oneself or others, or engage in valued recreational and social activities⁴. Early detection of impaired or declining function could support early rehabilitation or other interventions to preserve quality of life and sustain or improve function. Yet routine or longitudinal assessment of physical function in older cancer survivors has been slow to become standard of care.

The lack of focus on these assessments may be due in part to limitations of common measures of function, which include clinician-rated, patient-reported, and performance-based assessments. Clinician-rated performance status such as Eastern Cooperative Oncology Group (ECOG) Performance Status score is relatively coarse and prone to subjective biases and unreliability⁵. Patient-reported measures of physical function capture important aspects of patients’ experiences but are subject to patient recall and reporting biases and require infrastructure to collect, score, and interpret assessments. Performance-based measures of physical function such as the Short Physical Performance Battery (SPPB) can measure objective performance on standardized tasks but require time, resources, and space that limit their utility for repeated longitudinal monitoring. All three measures are limited in that they capture function at a single point in time and are often assessed only at pre-treatment, with limited potential for early detection of functional decline during or after cancer treatment.

Mobile devices provide new opportunities for estimation of physical functioning. These devices are capable of capturing behaviors like geographic mobility, physical activity patterns, and gait cadence (1) in the real world, which may better capture how people move around their typical environments during their routine daily activities; (2) over longer periods of time, which may be more illustrative of current functioning than a brief snapshot gathered at a single timepoint; and (3) passively, using devices that many patients already own and with minimal burden to participants. Smartphones, now owned by two-thirds of adults over age 65⁶, and consumer wearable devices such as wrist-worn activity trackers can accurately and continuously measure physical activity and mobility in older adults and patients with cancer^{7,8 9}. A number of small studies have examined correlations of simple metrics from wearable device data (e.g., step counts) and clinician- and patient-reported physical function in people with cancer¹⁰. These studies suggest that simple step count metrics from wearable devices are associated with traditional performance status and patient-reported functional outcomes, but none of these studies examined associations between wearable metrics and performance-based measures of physical function or focused specifically on older cancer survivors, who may be at elevated risk of impaired function and its consequences.

This field builds on work using research-grade accelerometers, metrics from which have been reliably associated with performance-based and patient-reported physical functioning in community-dwelling older adults but rarely examined in people with cancer^11,12,13. Further, their potential for longitudinal monitoring is limited by battery life and memory capacity¹⁴. Most of this work has also focused on the amount of physical activity in daily life. However, more recent studies have considered the clinical implications of more detailed information about activity patterns and behavior. For example, activity fragmentation, or activity bouts that are shorter and interrupted by rest bouts, is more common among older adults as well as cancer survivors¹⁵ and associated with mortality in older adults¹⁶. Gait speed measured during standardized walking tasks is also highly prognostic for morbidity and mortality among older cancer survivors¹⁷. Recent studies of older adults have reported that accelerometer-based activity fragmentation was associated with performance-based physical function¹⁸ and that free-living gait cadence from wrist accelerometers was a robust predictor of falls¹⁹, but these parameters have not been linked to functioning or outcomes after cancer treatment.

To our knowledge, no studies to date have linked consumer wearable activity monitor metrics, including step counts as well as activity fragmentation and gait cadence, with performance-based or patient-reported physical functioning in older cancer survivors. The goal of this pilot study was to examine the associations between consumer wearable device and smartphone sensor data and commonly used clinical measures of physical function in cancer survivors aged 65 and older. In exploratory analyses, we also examined consumer wearable device metrics, performance-based measures of physical function, and patient-reported physical function as predictors of fall risk.

Materials and Methods

Participants

Participants (n = 40) were recruited from two community research registries as well as oncology clinics. Eligibility criteria included: (1) had completed primary treatment (including surgery, chemotherapy, and/or radiation) within the previous five years for cancer of any type and stage except basal cell skin carcinoma; (2) aged 65 years or older; (3) were independently ambulatory (use of walking aids was acceptable); and (4) owned a smartphone capable of syncing the consumer wearable device and running the study applications.

Study Procedure

All participants provided informed consent, after which they were sent an electronic questionnaire that included demographic and clinical measure (including questions about cancer diagnosis and treatments and the presence of any comorbid conditions, listed in Table 1) as well as the Patient Reported Outcomes Measurement Information System (PROMIS) Physical Function Cancer measure²⁰ (PROMIS-PF, our primary measure of patient-reported physical function) as well as falls or near falls within the past month (“In the last month, have you fallen down? By falling down, we mean any fall, slip, or trip in which you lose your balance and land on the floor or ground at a lower level,” and “In the last month, have you slipped or tripped in which you lost your balance but did not fall down?”). Literature suggests that falls and near falls are highly associated²¹ and that near falls are a robust predictor of subsequent falls, and both are associated with injuries among older adults with cancer^22,23, so these were grouped into a single variable. Participants were then scheduled for an in-person assessment in a physical performance laboratory.

Table 1.

Baseline characteristics of the full sample and the sample with smartphone data.

	Full Sample (n = 40)	Sample with Smartphone Data (n = 25)
Mean age in years (SD)	73.35 (4.87)	73.24 (4.82)
Female, n (%)	20 (50%)	13 (52%)
White race, n (%)	39 (98%)	24 (96%)
Education, n (%)
High school or less	2 (5%)	2 (8%)
Some college/2-year degree	9 (23%)	5 (20%)
Bachelor’s degree	9 (23%)	7 (28%)
Graduate degree	19 (48%)	10 (40%)
Cancer type, n (%)
Prostate	17 (43%)	11 (44%)
Breast	9 (23%)	7 (28%)
Leukemia/lymphoma	5 (13%)	1 (4%)
Other	9 (23%)	6 (24%)
Cancer treatment received, n (%)
Surgery	28 (70%)	19 (76%)
Chemotherapy	15 (38%)	9 (36%)
Radiation	21 (53%)	15 (60%)
Immunotherapy	11 (28%)	7 (28%)
Hormonal therapy	8 (20%)	6 (24%)
Other	3 (8%)	2 (8%)
Mean months since cancer treatment (SD)	24.95 (22.02)	22.04 (19.68)
Presence of comorbidity, n (%)
Asthma, emphysema, or bronchitis	4 (10%)	2 (8%)
Arthritis or rheumatism	25 (63%)	15 (60%)
Diabetes	6 (15%)	3 (12%)
Digestive problems (e.g., ulcer, colitis)	3 (8%)	1 (4%)
Heart trouble (e.g., angina)	6 (15%)	4 (16%)
Kidney disease	2 (5%)	2 (8%)
Liver problems (e.g., cirrhosis)	2 (5%)	1 (4%)
Eastern Cooperative Oncology Group score of 0, n (%)	35 (88%)	21 (84%)
Mean PROMIS Physical Function score (SD)	49.30 (7.54)	50.84 (7.70)
Mean Short Physical Performance Battery score (SD)	9.57 (2.29)	9.56 (2.31)
Mean Timed Up and Go walk time in seconds (SD)	10.85 (2.85)	9.56 (2.31)
Mean daily step count (SD)	6428 (3282)	6531 (3346)

SD = standard deviation; PROMIS = Patient-Reported Outcomes Measurement Information System

In-person assessments took approximately one hour and were conducted by a physical therapist or cancer rehabilitation nurse with assistance from trained physical therapy students. Two primary measures of performance-based physical functioning were administered, and ECOG performance status²⁴ was also assessed. The SPPB assesses components of functional mobility including gait speed, time to complete five chair stands, and balance and yields a total score of 0–12 with higher scores indicating better physical function²⁵. The Timed Up-and-Go (TUG) is a measure of mobility, gait and balance, which assesses the time required to rise from a chair, walk three meters at comfortable pace, turn, and return to the seated position, with shorter times indicating better function²⁶

Following the performance-based testing, the AWARE and Fitbit apps were downloaded to participants’ Android or iOS smartphones, and they were provided a Fitbit Inspire 3 and asked to wear it continuously for four weeks except while charging. AWARE²⁷ was used to passively collect smartphone sensor data, including approximate location of the phone. This application runs in the background and stores raw data on device for encrypted transmission to our secure cloud-based research server. Fitbit was used to collect physical activity data, and minute-level step count data was accessed using the Fitbit Application Programming Interface (API). In addition to these activity and mobility measures of primary interest, AWARE also collected information about mode of transportation, smartphone battery level, metadata (e.g., timestamp and duration) of incoming and outgoing calls and texts, masked logged keyboard data, and screen on and off events, and Fitbit also collected heart rate and sleep data; these data are not reported.

Participants completed the same patient-reported measures of function and falls at the end of the study and received $50 in gift cards to compensate them for their time, in addition to the Fitbit device. The University of Pittsburgh and Duquesne University institutional review boards approved all study procedures.

Data Processing and Analyses

We used our Reproducible Analysis Pipeline for Data Streams (RAPIDS)²⁸ to compute day-level (24 hours from midnight to midnight) behavioral features from both AWARE and Fitbit data. These behavioral features were then averaged over all days on which sensor data was collected for at least eight hours, resulting in one summary feature per participant. For Fitbit data, we used the presence of heart rate data to infer continuous wear time and excluded days with fewer than eight hours of wear time. For the current analyses we focused on the following sensor features selected a priori to capture aspects of free-living activity amount, patterns, and gait speed: daily step count, peak gait cadence (maximum steps/minute), activity fragmentation (activity-to-sedentary transition probability, computed as the reciprocal of the mean active bout length, multiplied by 100), time spent at home (minutes), total distance traveled (m), and geographic mobility (radius of gyration, m; note the last two features were log transformed to normalize the distributions.).

To examine associations between sensor metrics and commonly used measures of physical function, we first assessed bivariate Pearson correlations between SPPB, TUG, and PROMIS-PF and the six sensor features listed above. We also conducted multivariable linear regression to determine whether sensor metrics were associated with physical function even when adjusting for important covariates including age and number of comorbidities. These regression analyses were conducted using SPSS Version 29. Finally, accidental falls are associated with serious injury among older adults, so we conducted exploratory analyses using the pROC package (version 1.18.4)²⁹ for R (version 4.2.3), to examine age, patient-reported physical functioning, performance-based physical functioning, and Fitbit sensor metrics as predictors of falls in the past month reported at the end of the four-week study; we conducted area under the receiver operating curve (ROC) analyses to determine the optimal cutoff threshold for the best predicting variables. We obtained 95% confidence intervals for area under the ROC curve (AUC) using 2,000 stratified bootstrap samples. To test the null hypothesis that AUC = 0.5, we conducted a series of Wilcoxon rank sum tests³⁰ and corrected for multiple comparisons by controlling the false discovery rate. We also conducted multivariable logistic regression to evaluate whether associations with falls were robust to adjustment for age and number of comorbidities.

Results

Participants’ characteristics are presented in Table 1. Overall, the sample was predominantly White and well-educated with a variety of cancer diagnoses and treatments. Most participants had received treatment within the past three years (70%). Participants endorsed a variety of comorbidities; number of comorbidities was collapsed into three groups: 0 = 23%, 1 = 55%, 2 or more = 23%. The majority of participants in our sample had an ECOG score of 0, or “fully active and able to carry on all pre-disease performance without restriction.”

At baseline, participants varied with regard to patient-reported physical function (PROMIS-PF mean 49.3, range 34.9–68.6) and performance-based physical function (SPPB mean 9.6, range 4–12; TUG mean 10.84 seconds, range 6.91–18.76), and these commonly used metrics were significantly correlated with each other (PROMIS-PF × SPPB: r(40) = 0.63, p < .001; PROMIS-PF × TUG: r(40) = −0.55, p < .001; SPPB × TUG: r(40) = −0.74, p < .001). On average, participants had 26 days of valid Fitbit data (range 12–29) and 15 days of valid phone sensor data (range 0–29). Due to technical issues with the sensor sampling frequency configuration for Android devices during AWARE data collection, 15 participants were missing phone data for the entire study, resulting in only 25 participants with sufficient phone data for analyses (subsample characteristics summarized in Table 1). The characteristics of the subsample of participants with smartphone data did not differ from those missing smartphone data. Given the amount of missing data, analyses with phone data features are based on a smaller sample size than intended and should be considered exploratory.

As shown in Table 2, Fitbit features related to greater amount and speed and to less fragmentation of walking in daily life were significantly related to better patient-reported (r’s 0.43–0.65) and performance-based (r’s 0.56–0.72) physical function. Smartphone sensor metrics related to greater geographic mobility were related to higher SPPB scores (r’s 0.40–0.42) but not patient-reported function or TUG scores.

Table 2.

Bivariate Pearson correlations between patient-reported and performance-based physical function and sensor metrics.

	PROMIS-PF	SPPB	TUG
Daily step count (n = 40)	.43, p = .006*	.56, p < .001*	−.61, p < .001*
Peak gait cadence (n = 40)	.65, p < .001*	.66, p < .001 *	−.72, < .001*
Activity fragmentation (n = 40)	−.52, p < .001*	−.69, p < .001*	.71, < .001*
Time at home (n = 25)	−.17, p = .43	−.38, p = .07	.12, p = .59
Total distance traveled (n = 25)	.18, p = .38	.42, p = .036*	−.17, p = .43
Geographic mobility (n = 25)	.28, p = .17	.40, p = .048*	−.19, p = .36

^*p < .05.

PROMIS-PF = Patient-Reported Outcomes Measurement Information System Physical Functioning; SPPB = Short Physical Performance Battery; TUG = Timed Up-and-Go.

We next conducted multivariable linear regression to examine how sensor metrics were associated with physical functioning, controlling for age and number of comorbid conditions. Table 3 shows coefficients from a series of models, the first including only age and number of comorbid conditions, followed by individual hierarchical linear regressions entering age and comorbidities on the first step and one of the sensor metrics on the second step.

Table 3.

Unstandardized (B) and standardized (β) estimates from linear regression analyses relating sensor metrics to patient-reported and performance-based physical function.

	PROMIS-PF			SPPB			TUG
	B (SE)	β	p	B (SE)	β	p	B (SE)	β	p
Covariates
Age(n = 40)	−.58 (.21)	−.37	.008*	−.17 (.06)	−.36	.011*	.17 (.08)	.30	.03*
Number of comorbid conditions (n = 40)	−4.60 (1.48)	−.41	.004*	−1.31 (.46)	−.39	.007*	2.08 (.55)	.50	<.001*
Daily step count (n = 40)	.00 (.00)	.21	.18	.00 (.00)	.39	.010*	.00 (.00)	−.46	<.001*
Peak gait cadence (n = 40)	.19 (.07)	.45	.007*	.07 (.02)	.50	.003*	−.09 (.02)	−.57	<.001*
Activity fragmentation (n = 40)	−18.62 (10.96)	−.27	.10	−11.55 (2.97)	−.55	<.001*	15.29 (3.40)	.58	<.001*
Time at home (n = 25)	.00 (.01)	.10	.60	.00 (.00)	−.10	.52	.00 (.00)	−.16	.35
Total distance traveled (n = 25)	−.13 (1.34)	−.02	.92	.47 (.33)	.21	.17	.15 (.48)	.05	.77
Geographic mobility (n = 25)	.52 (.81)	.11	.53	.27 (.20)	.20	.19	.03 (.29)	.01	.93

^*p < .05.

PROMIS-PF = Patient-Reported Outcomes Measurement Information System Physical Functioning; SPPB = Short Physical Performance Battery; TUG = Timed Up-and-Go.

After adjusting for age and comorbidities, faster peak gait cadence was associated with better patient-reported physical function (ΔR² = 0.12), but no other sensor metrics were significantly associated with function as assessed via the PROMIS Physical Function measure. In contrast, higher daily step count (ΔR² =0.12), faster peak gait cadence (ΔR² =0.15), and lower activity fragmentation (ΔR² =0.20) remained associated with higher SPPB scores after covariate adjustment. Similarly, higher daily step count (ΔR² =0.17), faster peak gait cadence (ΔR² = 0.19), and lower activity fragmentation (ΔR² = 0.22) were associated with shorter TUG walk time. Smartphone sensor features were not associated with any functional outcomes after covariate adjustment.

At the end of the study, 13/40 (32%) participants responded “yes” to a fall (n = 2) and/or near fall (n = 12) within the past month. We built a series of empirical ROC curves to evaluate the performance of classifiers using age, number of comorbid conditions, ECOG performance status, patient-reported physical function, performance-based physical function, or one of the three Fitbit metrics to predict past-month falls and/or near falls reported at the end of the study (see Table 4; note that our sample had insufficient variability in ECOG scores to estimate the optimal threshold value). Evaluated thresholds were the means of each pair of consecutive, unique values observed in the data. Optimal classification threshold was selected as the value that maximized Youden’s index. These preliminary findings suggest that a TUG score greater than 10.6, peak free-living gait cadence of less than 98.2 steps per minute, PROMIS-PF score less than 42.8, 2 or more comorbid conditions, or SPPB score of less than 8.5 are the best predictors of fall risk in this sample of older cancer survivors.

Table 4.

Exploratory receiver operating characteristic curve analyses classifying fall risk based on age, commonly used measures of physical function, and Fitbit metrics (n = 40).

	Direction	AUC	Bootstrapped 95% confidence interval	Adjusted p	Optimal threshold
Age	<	0.551	0.352–0.754	0.61	76.5
Number of comorbid conditions	<	0.721	0.563–0.866	0.03*	2 or more
Eastern Cooperative Oncology Group Performance Status score	<	0.464	0.370–0.578	0.61	n/a
PROMIS-PF	>	0.768	0.590–0.912	0.02*	42.8
SPPB	>	0.719	0.533–0.890	0.04*	8.5
TUG	<	0.802	0.628–0.940	0.02*	10.6
Daily step count	>	0.692	0.481–0.875	0.08	4920
Peak gait cadence	>	0.781	0.578–0.94	0.02*	98.2
Activity fragmentation	<	0.655	0.447–0.846	0.15	0.452

^*p < .05.

AUC = area under the curve. PROMIS-PF = Patient-Reported Outcomes Measurement Information System Physical Functioning; SPPB = Short Physical Performance Battery; TUG = Timed Up-and-Go. For Direction, < means the participant is classified as a fall/near fall “case” when the value of the predictor is greater than the threshold and a “control” otherwise, while > means the participant is classified as a “case” when the value of the predictor is lower than the threshold and a “control” otherwise.

We conducted exploratory logistic regression to examine how Fitbit metrics and commonly used physical function measures were related to fall group, after adjustment for age and comorbidities. Table 5 shows coefficients from a series of models, the first including only age and number of comorbid conditions, followed by individual hierarchical logistic regressions entering age and comorbidities on the first step and one of the functioning or sensor variables on the second step. Results from these multivariable analyses are consistent with the ROC analyses and indicate that peak gait cadence suggests a significantly protective effect and PROMIS-PF, SPPB, and TUG are marginally related to increased fall likelihood even after adjusting for age and comorbidities.

Table 5.

Exploratory logistic regression predicting risk of past month falls from physical function and Fitbit metrics, adjusted for age and comorbidities (n = 40).

	OR	95% confidence interval	p
Covariates
Age	1.02	0.87–1.19	0.86
Number of comorbid conditions	4.24	1.21–14.85	0.02*
PROMIS-PF	0.88	0.76–1.01	0.06
SPPB	0.67	0.44–1.01	0.06
TUG	1.41	0.98–2.04	0.07
Daily step count	1.00	1.00–1.00	0.64
Peak gait cadence	0.94	0.88–0.99	0.045*
Activity fragmentation	1.04	0.96–1.12	0.37

^*p < .05.

PROMIS-PF = Patient-Reported Outcomes Measurement Information System Physical Functioning; SPPB = Short Physical Performance Battery; TUG = Timed Up-and-Go.

Discussion

This study provides preliminary evidence that real-world data from consumer devices may be useful for estimating functional performance among older cancer survivors. Among participants aged 65 and older who had completed primary cancer treatment within the past five years, consumer wearable device data reflecting amount of activity, activity fragmentation, and peak gait cadence were significantly related to two commonly used performance-based measures of physical function, even after adjustment for age and comorbid conditions. On the other hand, smartphone sensor data reflecting geographic mobility were correlated with performance-based but not patient-reported function, and this association was not robust to covariate adjustment. In exploratory analyses, gait cadence from Fitbit data collected over four weeks was a significant predictor of fall risk even after adjusting for age and comorbidity.

This work supports the hypothesis that data that can be passively and remotely collected from consumer devices as patients perform routine daily activities could be useful for identifying individuals at risk for impaired physical function and falls and potentially for monitoring function longitudinally. Given the growing ubiquity of these devices, this approach to tracking functioning may be highly scalable, requiring no travel, time, or other effort from participants, and could be well-suited to remotely monitoring cancer survivors even over many years. Integrating sensor data into clinical workflows could help clinicians to identify patients who might benefit from rehabilitation, exercise, or other interventions and could also inform shared decision making around treatments to help optimize functioning and quality of life.

Accidental falls are highly prevalent among older adults with cancer, occurring in between 39% and 64% of community dwellers annually³¹—substantially higher than the 25% fall prevalence among the general community dwelling older adult population²¹. Unobtrusive passive monitoring for functional changes indicating fall risk could benefit such individuals. In the current pilot study, metrics related to gait cadence and activity fragmentation over a month of daily life were more strongly related to physical function and falls than simple daily step counts reflecting volume of physical activity. This finding extends work linking research grade accelerometer-based activity fragmentation and gait cadence to function and falls in older adults.^{18, 19, 32} Together with these other studies, this work suggests that free-living gait cadence and activity fragmentation capture clinically important patterns of activity that are not captured by step count or other measures of activity volume. The current study used consumer wearable devices rather than the research grade accelerometry used by these other studies and suggests that metrics of gait cadence and activity fragmentation can be computed from minute-level Fitbit step data and meaningfully linked to function. This study is also the first to examine these associations specifically in older cancer survivors.

Strengths of this study include the focus on older cancer survivors, a group at elevated risk for impaired physical function and for whom physical function is strongly prognostic; the assessment of both patient-reported and in lab performance-based physical function as well as behavior in daily life as evaluated using both a consumer wearable device and the participants’ own smartphone; and the consideration of more complex activity-related parameters than simple step counts. Limitations include the small, almost exclusively White sample and the requirement that participants own a smartphone; this may have biased the sample toward older cancer survivors with more socioeconomic advantage and digital literacy. It will be important for future work to include older cancer survivors from more marginalized backgrounds and to consider potential barriers to accessing the consumer technology used in this study in order to translate these findings to real-world clinical implementation. Another important limitation was the technical failure that resulted in missing smartphone sensor data for over one-third of the sample. To avoid similar data losses, future studies should check smartphone sensor data quality throughout the study to ensure that these passive data are being collected reliably. Participants were also followed for only four weeks, with only one performance-based physical assessment, which while sufficient for capturing activity patterns in daily life was likely not long enough to capture any significant decline or other change in function among these relatively healthy post-treatment survivors, nor was it sufficient for a more nuanced comparison with performance-based functional tests over time. We also averaged sensor metrics over the entire four-week period and did not consider potential variability or change in activity from day to day or week to week in the present analyses. Finally, our analyses of patient-reported falls or near falls³³ should be considered preliminary, given the small sample and relatively small number of falls limiting statistical power, especially given that we were unable to disentangle temporal effects of falls during the monthlong observation period impacting subsequent Fitbit-measured gait cadence. Future studies should examine these associations longitudinally and in larger and more diverse samples, which would allow for adjustment for additional risk factors for falls.

To our knowledge, this is the first study to link consumer wearable device measures of activity amount, fragmentation, and gait cadence to standardized measures of physical functioning in older cancer survivors. These preliminary results support the potential value of consumer sensor data in estimating and potentially remotely and longitudinally monitoring physical function during and after cancer treatment. Future studies should consider assessing older adults in active cancer treatment for longer than one month, which might permit detection of functional decline during chemotherapy or other treatments. Patients with advanced cancer or exhibiting functional impairment at diagnosis may also be important high-risk groups for longitudinal monitoring. Although physical function was more strongly related to the Fitbit metrics than to smartphone metrics, this may have been due to the reduced power for smartphone analyses, and replicating findings in a larger sample will be important.