Using Behavioral Analytics to Increase Exercise: A Randomized N-of-1 Study

Abstract

Introduction

This intervention study used mobile technologies to investigate whether those randomized to receive a personalized “activity fingerprint” (i.e., a one-time tailored message about personal predictors of exercise developed from 6 months of observational data) increased their physical activity levels relative to those not receiving the fingerprint.

Study design

A 12-month randomized intervention study.

Setting/participants

From 2014 to 2015, 79 intermittent exercisers had their daily physical activity assessed by accelerometry (Fitbit Flex) and daily stress experience, a potential predictor of exercise behavior, was assessed by smartphone.

Intervention

Data collected during the first 6 months of observation were used to develop a person-specific “activity fingerprint” (i.e., N-of-1) that was subsequently sent via email on a single occasion to randomized participants.

Main outcome measures

Pre–post changes in the percentage of days exercised were analyzed within and between control and intervention groups.

Results

The control group significantly decreased their proportion of days exercised (10.5% decrease, p<0.0001) following randomization. By contrast, the intervention group showed a nonsignificant decrease in the proportion of days exercised (4.0% decrease, p=0.14). Relative to the decrease observed in the control group, receipt of the activity fingerprint significantly increased the likelihood of exercising in the intervention group (6.5%, p=0.04).

Conclusions

This N-of-1 intervention study demonstrates that a one-time brief message conveying personalized exercise predictors had a beneficial effect on exercise behavior among urban adults.

INTRODUCTION

Despite the well-established health benefits of physical activity,^1–3 broad public adoption/maintenance of a physically active lifestyle remains elusive. Evidence suggests that less than 25% of U.S. adults meet physical activity guidelines.⁴ Although some interventions have been successful at increasing physical activity in the short term,^5–7 few interventions have succeeded in maintaining this health behavior.^5–7

Accumulating evidence suggests personally tailored interventions are more effective for increasing physical activity maintenance compared with one-size-fits-all approaches wherein an identical intervention is provided to all participants.^8–10 According to cognitive theory, people selectively pay attention to personalized information relevant to themselves and are more likely to retain it in their long-term memory.^11,12 This increases the likelihood of sustained change. As such, the emerging field of precision health emphasizes personalized strategies tailored to address each individual’s own behavioral, biological, and psychologic characteristics, as well as environmental exposures that influence health behaviors.^13,14 However, such interventions have been limited in that they rely upon perceived, participant-reported barriers to physical activity, and feedback is typically based on short observation periods (e.g., 1 week).^15–17 To enhance the efficacy of personalized physical activity interventions, tracking of objectively and subjectively measured physical activity patterns over sustained periods (e.g., weeks/months) may be warranted as a means for identifying features of daily life (e.g., stress) that influence whether or not individuals exercise on any given day.

With advances in mobile health technologies, the use of N-of-1 methodology^18,19—the study of the individual as a unique agent (within-subject, in contrast to the conventional between-subject approach) and the momentary level factors that predict physical activity—is now possible. Coupling wearable activity trackers (e.g., Fitbit) with ecologic momentary assessment (EMA) via smartphone, investigators can monitor physical activity while repeatedly collecting real-time exposure data on events that occur throughout the day, the experience of, and the behavioral, cognitive, and emotional responses to exercise. Combining these data sources allows for building personalized, observational, N-of-1 models elucidating factors that predispose each individual to exercise or not exercise on any given day (i.e., to develop a personalized “activity fingerprint”) and then to develop and deliver a personalized intervention. For example, through analyzing trends in data collected from multiple mobile health technology sources, one might find that for a given individual on days when they are stressed they are less likely to exercise. Providing this information in turn may cause the individual to reflect on why that is the case (e.g., self-reflection) and take action to change it on future stressful days. Along the lines of promising work by Bentley et al.,²⁰ this activity fingerprint intervention was developed to support reflection on personal data collected over a period of months and through this reflection start individuals on a path towards contemplative action and behavior change.

The activity fingerprint intervention was based on the Health Promotion Model, a validated model widely used in the promotion of physical activity.²¹ Its theoretic basis is focused on the multidimensional nature of individuals, in which there are interpersonal and environmental factors that interact to affect health behaviors. This Health Promotion Model posits that individual characteristics and experiences, feelings and knowledge about the desired health behavior, and immediate behavioral contingencies (e.g., response to immediate competing demands) influence healthy behaviors. According to the model, when an individual is made aware of the influencing factors/barriers to healthy behaviors, this stimulates cognitive processes that refine an individual’s perceived benefits, barriers, and self-efficacy towards these health-promoting behaviors, and in turn facilitates engagement in that behavior.

The aim of the present study is to use mobile health technologies to investigate whether those randomized to receive a personalized activity fingerprint increased their physical activity levels compared with a control group. The primary hypothesis is that those who receive the personalized activity fingerprint will increase the proportion of days they exercise following receipt of the fingerprint relative to the control group.

METHODS

This was a 12-month observational cohort study that entailed intensive longitudinal data collection for the entire study duration and was designed to examine the bidirectional association between stress and exercise.²² After 6 months of observation, as an exploratory aim, participants were randomized to a control or intervention group, the latter receiving a one-time communication via email regarding personalized factors that predicted whether or not they exercised on any given day (the activity fingerprint). The study was conducted at Columbia University Medical Center, an academic medical center in New York City. Data were collected from January 2014 to May 2015. Eligible participants were enrolled and then initiated daily accelerometer-based monitoring of physical activity and smartphone-based EMA reports of stress and daily exercise (Appendix Figure 1). Data collected over the first 6 months of observation were used to develop an individualized activity fingerprint for each participant, identifying the individual factors that predicted whether they exercised or not on any given day. All participants received a one-time report summarizing their physical activity and stress levels during the prior 6 months. For approximately half of the participants, randomly assigned, the one-time report also included their activity fingerprint—a brief statement describing one to three factors (e.g., anticipated stress level on a given day) that either increased or decreased the likelihood of their exercising. The remaining participants’ reports did not include the activity fingerprint (Figure 1). The study was approved by Columbia University’s IRB; all participants provided informed consent. Access to the study dataset and information about the study’s execution and materials is publicly available at https://osf.io/kmszn.

Figure 1.

CONSORT diagram.

Study Population

Postings throughout the Columbia University Medical Center campus were used to recruit a convenience sample of 79 individuals. Eligibility criteria were: English speaking, aged ≥18 years, self-reported intermittent engagement in exercise, daily access to computer with Internet, and use/ownership of an iPhone/Android phone. Excluded were individuals who had previously been told by a healthcare professional to restrict physical activity, were deemed unable to comply with the protocol for 12 months, or had serious medical comorbidity.

Measures

This study applied a triangulation approach²³ for measuring physical activity. Specifically, self-reported physical activity was used in addition to objective physical activity data because of concerns that accelerometers are not capable of capturing some of the more common forms of aerobic exercise. Research- and commercial-grade (such as those made by Fitbit) accelerometers have poor accuracy for the measurement of cycling and cannot be worn while swimming because of not being waterproof.^24–26 Thus, self-report and objective physical activity data were concurrently collected. This triangulation approach was chosen with the goal of providing a more complete picture of participants’ physical activity.

Objective physical activity data were continuously collected using a wrist-based model of the Fitbit (Fitbit Flex; www.fitbit.com/company). The Fitbit Flex is a microelectromechanical triaxial accelerometer that tracks the wearer’s physical activity including steps, intensity of activity (sedentary, light, moderate, or vigorous), and energy expenditure. The Fitbit Flex is a valid and reliable device for measuring physical activity in adults.²⁷ Participants were not provided access to the account that was linked to their device. Thus, they did not have access to any of their objectively assessed physical activity data. Details concerning accelerometer data processing are described in the Appendix Methods. Following accepted standards,²⁷ a 30-minute period of exercise was objectively defined as any 30-minute period within which ≥24 minutes were classified by the Fitbit as moderate or vigorous intensity activity. Each valid day was then classified as an exercise or non-exercise day.

An electronic diary displayed on the participant’s own iPhone/Android phone browser was used to capture momentary and summary aspects of stress daily for the 12-months of observation. Participants were asked to complete five EMA surveys each day (morning, evening, and three random mid-day surveys) for the entire 12-month study period. Each morning, participants responded to questions on a 0 (Not at all) to 10 (Extremely) scale asking them, (1) How stressful do you expect today to be? and (2) How likely are you to exercise today? Similarly, each evening participants responded to questions asking them, (1) Overall, how stressful was your day? and (2) Overall, how stressful do you think tomorrow will be?, again answered on a 0 (Not at all) to 10 (Extremely) scale. They were then asked whether or not they had exercised for ≥30 minutes that day (Yes/No).

In addition, the system was programmed to prompt the participant via text message/email three random times over their preset hours of wakefulness. Assessments included questions concerning key sources of stress (e.g., work, home, financial, interpersonal, time/scheduling pressure, daily hassle, or general/other, with the participant checking all that apply) and stress appraisal, using the four-item Perceived Stress Scale.

Intervention

This study was a randomized controlled intervention nested within a 12-month observational cohort study. Data collected during the initial 6-month period of observation on anticipated and perceived stress (e.g., How stressful do you expect today to be? Overall, how stressful was your day?), mid-day level of stress, self-determination of exercise (e.g., How likely are you to exercise today?), and weekly temporal patterns of exercise behavior (e.g., more likely to exercise on weekday versus weekend), along with 1- and 2-day lagged versions of these measures, were used to generate a unique activity fingerprint for each individual (i.e., up to three independent predictors of exercising for each participant). At 6 months, all participants received a descriptive report via email summarizing their daily stress reports and daily minutes of exercise (Appendix Figure 2); approximately half of participants, randomly assigned, also received a two to four sentence summary of their personalized activity fingerprint (Appendix Figure 3). The set of exercise predictors considered for inclusion in the activity fingerprint are listed in Table 1. The most common components of the activity fingerprint reflected self-determination, perceived stress, and weekly temporal patterns. Previous studies have documented that self-determination,^6,28–30 perceived stress,³¹ and weekday/weekend differences are factors associated with exercise participation.^32–34 Additional details concerning the generation of the activity fingerprints and the theoretic basis of the intervention are provided in the Appendix Methods.

Table 1.

Variables Considered for Inclusion in Personalized Fingerprints

Concepts	Measurements
Self-determination	Morning rating of likelihood of exercising todayb
Perceived stress	Last night’s summary rating of actual stress for yesterdaya
	Last night’s rating of expected stress todaya
	Morning rating of expected stress todayb
	Midday rating of stressb
	Evening summary rating of actual stress today
Temporal patterns	Day of the week
	Weekend versus weekday
Exercise	Last night’s report of whether exercised yesterday (self-report)a
	Whether had 30-minute bout of exercise yesterday (by actigraphy)a

^aAlso considered was this variable assessed the day before yesterday and the day before that.

^bAlso considered was this variable assessed yesterday, the day before yesterday, and the day before that.

Statistical Analysis

All randomized participants (n=73) were included in an intent-to-treat analysis using all days with available data for the exercise measure being analyzed; days with missing exercise data were assumed to be missing at random. The primary outcome was the daily measure of exercise (Yes/No); the analysis yielded group specific estimates of the probability of exercise pre- and post-fingerprint. An exercise day was defined as either an objectively assessed bout of exercise ≥30 min or self-reported exercise ≥30 min via EMA. Because of concerns of a Hawthorne effect from accelerometer wear/EMA responses during the initial weeks after study enrollment and concerns that the hypothesized benefit of the fingerprint might not be sustained (i.e., would dissipate) over time, investigators chose not to compare the entire 6 months pre-fingerprint to the 6 months post-fingerprint. Instead, investigators decided a priori to examine the 3-month period (91 days) following the 6-month communication, and compare this to the 3-month period immediately preceding the communication.

To test the overall effect of the activity fingerprint intervention, a Poisson generalized estimating equation (GEE) model was estimated predicting the daily binary exercise outcome from group (fingerprint, yes/no), period (pre- versus post- the date the fingerprint was sent), and group ×time as predictors. A compound symmetry error structure was assumed, after finding that it provided a better fit than an autoregressive (AR) model error structure. From model estimates, estimates of relative risk were obtained between groups within period and between periods within group and were able to calculate corresponding estimates of the percentage days exercised for the fingerprint and no-fingerprint groups, and the differential change between groups (primary hypothesis; described further in Appendix Methods).

As an exploratory analysis, in order to explore the extent of individual differences in the pre–post change in the proportion of days exercised, each participant with usable data pre- and post-intervention (n=72) was treated as an N-of-1 study and estimated a Poisson regression with allowance for autocorrelated residuals obtaining estimates of the relative risk (RR; and its 95% CI) of daily exercising post- versus pre-fingerprint. Using standard meta-analysis formulae, within-group heterogeneity in the estimates of change was tested and if significant, the variance of this heterogeneity was estimated.³⁵

As a sensitivity analysis, the intent-to-treat GEE analysis was repeated examining objectively measured exercise and self-reported exercise as separate outcome variables (n=73 for both outcome variables). All statistical analyses were performed using SAS, version 9.4. The α level was set at p<0.05.

RESULTS

All randomized participants (n=73) provided usable data for the 3 months pre-fingerprint and all but one provided usable data for the 3 months post-fingerprint. During the 12-month study period, the mean number of days with a morning EMA report, evening EMA report, or ≥1 midday EMA report completed were 213.3 (SD=79.0), 226.4 (SD=81.7), and 350.5 (SD=33.2) days, respectively. The mean number of valid accelerometer wear days was 242.6 (SD=76.7 days). Among the 73 participants, mean age was 31.9 (SD=9.6 years), 41.1% were male, 58.9% were single, and 29.2% were Hispanic. There were no significant differences between those who received their activity fingerprint (n=39) and those who did not (n=34) in demographic characteristics (Appendix Table 1). There were also no significant differences between the two groups in valid days of exercise assessment (≥10 hours of accelerometer wear and EMA response) pre- or post-fingerprint.

Components of the personalized exercise prediction models (fingerprints) for those in the intervention group and their respective changes in exercise rates are summarized in Appendix Table 2. The prediction models built from each participant’s EMA and Fitbit data included the following components:

• morning prediction of exercise/exercise determination (n=12, 30.8%; e.g., “your morning expectation of exercising has an effect on the likelihood of actually exercising…”);

• stress anticipation (n=5, 12.8%; e.g., “your evening expectation of how stressful your tomorrow will be, has an effect on the likelihood of exercising…”);

• EMA-based perceived stress score during the day (n=8, 20.5%; e.g., “how stressful your day has been has an effect on the likelihood of having exercised”);

• temporal period referenced–today (n=10, 25.6%; e.g., “…today, …exercise [today]”);

• temporal period referenced–multiple days: today and 1, 2, or 3 days prior (n=21, 53.8%; e.g., “if you exercised two days ago, you are more likely to exercise today”); and

• weekend/weekday patterns (n=9, 23.1%; e.g., “you are more likely to exercise on weekdays rather than weekends”).

There was very little difference between those who did versus did not receive the fingerprint in the percentage of days exercised (≅44.0%) during the 3 months prior to randomization (Table 2). The group that received their fingerprints exhibited a slight, nonsignificant reduction in the percentage of days exercised during the subsequent 3 months (4.0%, RR=0.91, p=0.14). The control group—those randomized to not receive their fingerprint—exhibited a larger, statistically significant reduction in the percentage of days exercised (10.5%, RR=0.76, p<0.0001). The decrease in the intervention group was significantly less than that in the control group (6.5%, RR=1.20, p=0.04).

Table 2.

Estimates of the Mean Percentage of Days Exercised During the 3 Months Pre- and Post-fingerprint, by Group

Group	Pre-fingerprint	Post-fingerprint	Change	p-value
	Mean (95% CI)	Mean (95% CI)	RR (95% CI)
Control group (no fingerprint, N=34)	42.9% (36.2, 50.9)	32.4% (26.2, 40.1)	0.76 (0.67, 0.85)	<0.0001
Intervention group (fingerprint, N=39)	44.0% (37.9, 51.1)	40.0% (33.5, 47.6)	0.91 (0.80, 1.03)	0.14
Group difference, RR	1.02 (0.82, 1.28)	1.23 (0.93, 1.62)	1.20 (1.01, 1.43)	0.04
p-value	0.83	0.14

Note: Boldface indicates statistical significance (p<0.05).

RR, relative risk

Exercise and non-exercise days over the course of the study are displayed at the individual level in Appendix Figure 4. The forest plot (Figure 2) illustrates individual level change (RR from pre- to post-fingerprint) in exercise rate; RR ranged from 0.06 (94% decrease) to 1.90 (90% increase) in the control group and from −0.22 (78% decrease) to 2.48 (148% increase) in the intervention group. The average RR (redline) in the control group was shifted more to the left than the average RR (redline) in the intervention group, consistent with the greater reduction in the proportion of exercise days in the control group. In absolute terms, none of the control group participants exhibited an increase of ≥10% in the percentage of days exercised, but 14 (41.2%) exhibited a decrease of ≥10% in the percentage of days exercised. In the intervention group, 16 participants (42.1%) showed a decrease in the percentage of days exercised of ≥10%, but seven (18.4%) showed an increase of ≥10%. Within each group, the variability among participants in the pre- to post-fingerprint change in the percentage of days exercised (Q statistic) was statistically significant (p<0.0001).

Figure 2.

Forest plot of individual changes in percent of valid days exercised with 95% CI. Red line shows mean change for group.

Notes: Post- vs pre-fingerprint prevalence ratio of >30 minutes of moderate to vigorous physical activity, with 95% CI.

In sensitivity analyses where exercise was based on EMA self-reports only, the control group exhibited an 8.2% decrease (RR=0.77, p=0.0003) in the percentage of days exercised whereas the intervention group had no change (0.0%, RR=1.00, p=0.99); the differential change of 8.2% was statistically significant (RR=1.30, p=0.02). When exercise was assessed only by accelerometry, the control group exhibited a 10.7% decrease (RR=0.72, p<0.0001) in the percentage of days exercised whereas the intervention group had a 6.3% decrease (RR=0.82, p=0.01); the differential change of 4.4% was in the hypothesized direction, but not statistically significant (RR=1.14, p=0.24).

DISCUSSION

In this randomized controlled intervention, it was tested if personalized feedback (i.e., an activity fingerprint) would affect daily physical activity. N-of-1 observational predictor models of exercise (the activity fingerprint) developed using the first 6-months of daily data were delivered to approximately half of the participants, randomly assigned, with remaining participants receiving only a descriptive stress/exercise report. This was done to test the Health Promotion Model of behavior change; that individual characteristics and experiences, feelings and knowledge about the desired health behavior will alter that behavior. These findings indicate that a single dose of a brief, personalized message conveying empirically derived factors based on 6 months of behavioral data had a beneficial effect on exercise behavior among those who received the activity fingerprint (intervention group) relative to those who received only a stress/exercise descriptive report (control group). Specifically, the intervention group maintained pre-intervention exercise behaviors (i.e., there was no significant change from pre- to post-intervention) relative to a control group whom exhibited a decline in exercise from pre- to post-intervention.

Prior research shows that the propensity to engage in intentional physical activity or exercise is influenced by a number of factors that can vary across individuals, indicating the potential for personalized approaches or precision health for increasing exercise. The current study used mobile health technologies to build N-of-1 models for predicting the likelihood of exercise and then examined the impact of providing a brief summary of the individualized model on a per person basis. In 2015, NIH launched the Precision Medicine Initiative to encourage scientists and medical providers to use creative new approaches to assess behavioral and environmental parameters to achieve the goal of precision health.¹³ With the pervasive use of smartphones (64% of U.S. adults own a smartphone)³⁶ and inexpensive wearable devices (e.g., Fitbit; 49% of U.S. adults own a wearable fitness tracker³⁷), a unique opportunity exists to first collect ecologically valid and objectively measured data over long periods of time with low participant/patient burden, and then use these data to inform practices and alter targeted health behaviors.³⁸

Several studies have demonstrated the utility of combining data collected over extended periods from multiple consumer devices/sensors as a tool to promote behavior change. Bentley and colleagues²⁰ developed a Health Mashups mobile application that integrated automatically sensed smartphone (weather, GPS location), wearable/health device (physical activity, sleep, weight), and smartphone logged (dietary intake, mood, pain) data continuously in real-time and delivered informative observations (“On days when you sleep more, you get more exercise”) daily to increase self-awareness of health behavior engagement and facilitate self-reflection. Improvements in mood, weight loss, and well-being were observed after a 90-day trial; although absence of a control group precluded conclusions that delivery of the informative observations contributed to the reported changes. Similarly, Choe et al.³⁹ developed a SleepTight mobile application that integrated sleep, dietary intake, physical activity, caffeine/alcohol consumption, and medication usage data in real-time and provided information to participants on how these factors influenced sleep hygiene. Although no formal quantitative sleep data were analyzed, qualitative data suggested that information provided from the application elicited self-reflection on sleep habits and the activities/behaviors that influence such habits; an important tenet of behavior change.⁴⁰ The present study extends upon previous findings by demonstrating within the context of a randomized controlled intervention that personalized feedback derived from long-term data collected via multiple consumer devices elicited improvements in a health behavior (exercise) relative to controls.

This study demonstrates a practical use of mobile health technology to continuously capture momentary health behaviors (exercise) and daily life features (stress) for extended periods (6 months) in order to develop a personalized intervention. Previous studies that have developed personalized physical activity interventions have been limited by short-term objective measurements (e.g., 1 week).^5,8,9,41 Further, none of these used N-of-1 modeling to identify personalized predictors. As conventional accelerometers have to be retrieved to upload data and have limited storage capacity, the wireless capability of the authors’ chosen device permitted capture of real-time data over an extended period (1 year), thus allowing for assessment of long-term exercise habits and capture of sufficient data to enable modeling of exercise predictors. Smartphone-based EMA permitted assessment of one potential barrier to physical activity (stress) on a daily basis. Using predictive statistical methods, actual factors (versus perceived) that predispose one to exercise/not exercise were therefore able to be ascertained. The finding that a single dose of a personalized message describing factors that predispose one to exercise/not exercise had a statistically significant effect on exercise behavior in the intervention group relative to controls is promising and supports the technical and operational feasibility of long-term use of smartphones and wearable devices for other similar N-of-1 trials.⁴² However, caution is warranted when interpreting these results given that within-group analyses showed that the intervention group only maintained pre-intervention exercise behaviors, but did not exhibit the hoped for increase following intervention delivery.

In the present study, in order to develop the individualized predictor models, 6 months of observational data was collected during the pre-intervention phase on self-determination of exercise (i.e., How likely are you to exercise today?), anticipated and perceived stress (i.e., How stressful do you expect today to be? Overall, how stressful was your day?), level and sources of stress at a given moment, and weekly temporal patterns of exercise behavior (i.e., “more likely to exercise on weekends rather than weekdays”). These factors were used because previous studies have documented that self-determination,^6,28–30 perceived stress,³¹ and weekday/weekend differences are factors associated with exercise participation.^32–34 For some participants, the individual predictor models developed based on these factors were likely useful to maintain exercise. For example, activity fingerprints containing information on how anticipated stress influenced exercise habits were useful as 80.0% of participants (four of five participants) who received this fingerprint had a favorable outcome. However, for others the individual predictor models may not have been useful. For example, only 44.4% of participants (four of nine participants) had a favorable outcome among those who received a fingerprint containing information on weekday/weekend differences in exercise habits. Future studies should examine other factors demonstrated to influence daily exercise behavior including environmental factors (e.g., weather), biobehavioral factors (e.g., sleep), and psychologic factors (state versions of perceived autonomy, competence, self-efficacy, and barriers).

Limitations

Several limitations should be noted when interpreting the study findings. First, study generalizability may be limited as participants were recruited from a single urban university. Second, there were moderate amounts of missing accelerometer/EMA data; however, there were no differences in the number of days of missing data between intervention and control groups. Third, a limitation of employing mobile health monitoring methods is that they themselves may influence behavior, rather than just record it. As such, observed decreases in exercise within both the intervention and control groups may be reflective of the dissipation of a Hawthorne effect. Fourth, data on whether the email delivering the intervention was opened/read was not collected. Finally, results of person-specific Poisson regression models indicate that the residuals exhibited a small serial autocorrelation (average person-specific correlation was 0.09). GEE analyses were unable to adjust for autocorrelation while also allowing for the much more substantial between-person differences in the probability of exercising. Given the robustness of GEE estimates and SEs to misspecification of the correlation structure, it is unlikely this limitation substantially affected results.

CONCLUSIONS

This study illustrates the feasibility of a novel methodology that utilized a large volume of ecologically valid data collected continuously over an extended period of time (6 months) to comprise a personalized physical activity intervention. Importantly, this study demonstrates a novel use of mobile health technologies (smartphone, physical activity trackers) to develop personalized interventions. The finding that a one-time brief message conveying personalized predictors of exercise had a beneficial effect on exercise behavior among adults relative to controls provides an example of how mobile health technologies may be a viable tool for precision health.