In-hospital physical activity measured with a new Bosch accelerometer sensor system

Abstract

Extended hospital stays and readmissions are associated with high costs and burden of disease. Effective ways to predict patient outcomes, select those at risk of adverse outcomes, and to prevent these accordingly, are lacking. Physical activity measured by accelerometry is a promising surrogate parameter of health, but it is mainly evaluated in highly selected patient groups and with few algorithms. We used a new wrist-worn Bosch sensor platform to measure various characteristics of physical activity in 58 patients of various ages treated on internal medicine wards over a period of 30 days, with a goal of developing algorithms to predict duration of hospitalization and 30-day readmission. New algorithms are required to increase the accuracy of the predictions and to understand the significance of physical activity in hospitalized patients. To facilitate research in this area, and to allow the public to explore these issues, all data generated in the study are published here as a public dataset.

Background

Effective ways to predict patient outcomes, select those at risk of adverse outcomes, and to prevent these accordingly, are lacking. Traditional risk factors associated with readmission are admission diagnosis, age, comorbidities, social factors, polypharmacy, among others [1]; however, these factors show poor predictive performance [2]. In the context of limited healthcare resources, accelerometers theoretically offer the possibility to observe mobility and physical activity (PA) as a promising surrogate parameter of health, at least in surgical patients [3].

Therefore, the aim of this study is to characterize PA in a mixed internal medicine in-hospital population, including a variety of frequent-admission diagnoses (acute infections, exazerbations of chronic obstructive lung disease (COPD), heart failure and fluid overload) and younger patients. Furthermore, we aimed to test this approach in studying the associations of different measures of PA with length of stay and 30 day re-admission status as traditional outcome measurements, and test the additional value of body position measurements using a 3G wrist-worn sensor system.

Methods

The measurements in this prospective observational study took place in 3 general internal medical wards with a total of 100 beds, and a 30-bed oncology ward in a 1100-bed referral center at the Robert-Bosch hospital in Stuttgart, Germany, over a 20-day period. The study was conducted with approval of the Ethics Commission at the Medical Faculty of the Eberhard Karls University and at the University Hospital Tübingen, Germany (507/2016BO1). For data protection reasons, no exact recording time is given. However, in order to be able to classify the data in its topicality, the information is provided that the data was collected in 2016.

Eligible subjects were at least 18 years old, able to walk 3 meters with or without an assistive device, able to provide informed consent, had no medical contraindication against wearing a watch on one wrist, and were admitted as a result of diagnosis of the following acute medical conditions: fluid overload in patients with CKD or heart failure, acute infections, COPD exacerbations, or unplanned hospitalization owing to progression of malignancy or infectious complications of oncologic treatment. Patients with amputations, and those who were only wheelchair-mobile or under bed-rest were excluded. Patients were also required to have an anticipated length of stay of at least 48 hours. Patients were monitored for a maximum of 10 days, if not discharged earlier.

Clinical assessment

Data were extracted from the Hospital Information System (HIS). Demographic data (age, gender) and information for length of hospital stay, readmission history, and diagnosis were obtained from discharge letters; additionally, the Charlson Comorbidity Index (CCI) scores were obtained at discharge and at 30 days post-discharge [4] . Elective readmissions were censored from the analyses.   

Clinical variables included laboratory data (C-reactive protein (CRP), creatinine, white blood cell count, and hemoglobin concentration), heart rate, body temperature, blood pressure, blood oxygen saturation, and number of prescribed medications.

On inclusion, subjects were surveyed using a questionnaire with following items: age, sex, ethnicity, number of falls within the last 12 months, last hospital stay, height, weight (calculated BMI), dominant hand, smoker status, average consumption of alcoholic beverages, use of walking aids indoors and outdoors (rollator, crane, wheelchair); housing status and nursing aids (autonomously, ambulatory nursing aid, nursing home), and a German-translated short IPAQ questionnaire. A timed “up&go” test was performed after the wrist sensor was applied for the first time [5]. 

Falls and mobility status were assessed on daily visits. Mobility was scored by the patient using a visual analog scale with the question “compared with your normal average physical activity at home, how would you score yesterday’s physical activity on a scale from zero (no activity at all) to ten (normal activity)?”

Accelerometry

The device, developed at Bosch Sensor Tech (as in Pozaic [6]) provides the following measurements: Acceleration (acc), Angular rate (gyro), Light, Temperature, Humidity, and Air pressure. Only raw sensor signals were collected. In order to ensure a minimum length of data per record, the following criteria had to be met: Recorded data of at least 8 hours, and time in which the device was worn greater than 80 percent of the total recorded duration .

Sampling rates were 100 Hz for the inertial sensor (acc + gyro) and 1 Hz for all other sensors. Patients were encouraged to wear the wristband at all times, except when taking a shower or bath.

Data Description

Each patient included in the study is represented by a folder (R401 - R458).

Data structure

Each folder has the following structure:

* R4xx
*  |- export
*       |- examineeInfo_publicReduced.json
*       |- measurement_1
*            |- sensorData.csv
*            |- measurementInfo_publicReduced.json
*       |- ...
*       |- measurement_k
*            |- sensorData.csv
*            |- measurementInfo_publicReduced.json
*
* For practical reasons the dataset is split into archives including up to 
* ten patients each. 

Files

• examineeInfo_publicReduced.json: Provides information such as age, gender, etc. for the patient in json format.
• measurementInfo_publicReduced.json: Provides information such as vital and lab parameters, weight etc. sampled at the day of the measurementInfo_publicReduced.
• sensorData.csv: Contains the sensor signals measured in each patient. Data is in standard comma separated (csv) format. The following signals are contained:
  • time: in ms after measurement start
  • acc_1-3: Three axial acceleration signal
  • gyro_1-3: Three Axial rotation rate signal
  • hum: humidity signal in %
  • light: light in J
  • pres: pressure signal in kg·m−1·s−2
  • temp: temperature in degree centigrade

Usage Notes

The motivation for collecting this dataset was to investigate correlations between physical activity and health condition, in particular within a hospital setting. To this end, data of in-patients was collected via wrist worn sensors.

Sensor data

Sensor data is available within the sensorData.csv files and include acceleration, gyroscope and some environmental data like humidity, air pressure or temperature. For some patients several measurements on different days of their hospital stay were collected.

The quality of the data is a compromise between providing the data as raw as possible and making the data as easily usable as possible. Hence, this data is mostly in raw format, however some preprocessing was done in order to address some sensor related and technical obstacles. The sensor data does not contain any form of derived "physical activities" but signals from the individual sensors. This enables researchers to investigate new forms of activity recognition algorithms (ARA) suitable and specific to the evaluation of a patient’s health condition, rather than to rely on generic activity measures such as e.g. step counts or other standard activities available in most consumer products.

Data on health conditions

Data on health conditions was collected on two levels. For one the general patients (time invariant at least during the hospital visit) condition, such as gender, age, length of the total hospital visit or normal housing condition. This information is provided within the examineeInfo_publicReduced.json file.

The second level was on data that varies between measurements, such as body temperature, blood pressure or some standard laboratory values collected at the same day the sensor data was recorded. This information is provided within the measurementInfo_publicReduced.json file. By this approach we want to provide the possibility to investigate both, the correlation between acute changes of activity and health condition (on an individual level) as well as the correlation between changes of activity and health condition in general.

Limitations

• Data quality and availability is variable. For some patients there may be few valid records available, or even no records.
• The item “Readmission status” reports only readmissions to the index hospital, therefore possibly underestimating the rate of readmissions.
• We employed one single wrist-worn sensor for detecting movements and posture.

Example usage

Usage of the dataset highly depends on the question in hand. However, a study might involve processing the dataset in the following way:

• Load sensor data (sensorData.csv)
• Apply some kind of activity recognition algorithms (ARA)
• Use ARA on sensorData.csv to generate a feature of interest
• Compare these features in their quality and / or in regard to one or more outcome measures with each other or to some measures of health conditions or outcomes either on:
  • a measurement level (measurementInfo_publicReduced.json)
  • a patient level (examineeInfo_publicReduced.json)
  • or a combination of both

Existing studies employing this dataset

In our personal analysis of the data we used several different standard and self-developed ARAs to convert the sensor data into different features of "physical activity" and investigated to what extend these features correlate to the physical conditions of patients, both on individual and general level. We expect results to this study to be published in the near future.

Acknowledgements

This research was funded by the Robert Bosch Foundation and Bosch Healthcare Solutions GmbH. We would like to thank Bosch Healthcare Solutions GmbH for providing sensors and technical support for this study.

Conflicts of Interest

N.S. is an employee of Bosch Healthcare Solutions GmbH, which provided the sensors and technical support for the study. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the analyses, or in the decision to publish the results or data set.

References

1. C. R. Carpenter et al., "Risk factors and screening instruments to predict adverse outcomes for undifferentiated older emergency department patients: a systematic review and meta-analysis," (in eng), Acad Emerg Med, vol. 22, no. 1, pp. 1-21, Jan 2015.
2. D. Kansagara et al., "Risk prediction models for hospital readmission: a systematic review," (in eng), Jama, vol. 306, no. 15, pp. 1688-98, Oct 19 2011.
3. A. Abeles, R. M. Kwasnicki, C. Pettengell, J. Murphy, and A. Darzi, "The relationship between physical activity and post-operative length of hospital stay: A systematic review," (in eng), Int J Surg, Review vol. 44, pp. 295-302, Aug 2017.