key: cord-0994932-04gbg8eu
authors: Halfwerk, Frank R.; van Haaren, Jeroen H. L.; Klaassen, Randy; van Delden, Robby W.; Veltink, Peter H.; Grandjean, Jan G.
title: Objective Quantification of In-Hospital Patient Mobilization after Cardiac Surgery Using Accelerometers: Selection, Use, and Analysis
date: 2021-03-11
journal: Sensors (Basel)
DOI: 10.3390/s21061979
sha: 4fc9d2c596be7ef3fc9fcedc4d16748eae26127c
doc_id: 994932
cord_uid: 04gbg8eu

Cardiac surgery patients infrequently mobilize during their hospital stay. It is unclear for patients why mobilization is important, and exact progress of mobilization activities is not available. The aim of this study was to select and evaluate accelerometers for objective qualification of in-hospital mobilization after cardiac surgery. Six static and dynamic patient activities were defined to measure patient mobilization during the postoperative hospital stay. Device requirements were formulated, and the available devices reviewed. A triaxial accelerometer (AX3, Axivity) was selected for a clinical pilot in a heart surgery ward and placed on both the upper arm and upper leg. An artificial neural network algorithm was applied to classify lying in bed, sitting in a chair, standing, walking, cycling on an exercise bike, and walking the stairs. The primary endpoint was the daily amount of each activity performed between 7 a.m. and 11 p.m. The secondary endpoints were length of intensive care unit stay and surgical ward stay. A subgroup analysis for male and female patients was planned. In total, 29 patients were classified after cardiac surgery with an intensive care unit stay of 1 (1 to 2) night and surgical ward stay of 5 (3 to 6) nights. Patients spent 41 (20 to 62) min less time in bed for each consecutive hospital day, as determined by a mixed-model analysis (p < 0.001). Standing, walking, and walking the stairs increased during the hospital stay. No differences between men (n = 22) and women (n = 7) were observed for all endpoints in this study. The approach presented in this study is applicable for measuring all six activities and for monitoring postoperative recovery of cardiac surgery patients. A next step is to provide feedback to patients and healthcare professionals, to speed up recovery.

Cardiac surgery consists mainly of coronary artery bypass grafting (CABG) followed by valve surgery [1] . The hospital stay for these patients is 5 to 7 days after surgery. Currently, automated monitoring of the mobilization process in the period surrounding the surgery is sparsely applied in practice. Table 1 . Requirements for a device to classify in-patient physical activities after cardiac surgery, ranked from highest to lowest importance.

Clinical Requirements

Classification of various activities Static: Lying, sitting, standing Dynamic: Walking, cycling, walking the stairs 1.

No burden during the hospital stay and no hindrance in mobilization 2.

Sampling rate between 60 and 100 Hz 2.

Minimal soft tissue deformation 3.

Battery life for measuring up to 7 consecutive days 3. Device placement away from wound area 4 . Local data storage up to 7 consecutive days 4.

Meeting regulations regarding patient privacy and integrity 5.

Data extraction from the device 5.

No impediment in daily nursing care 6.

Device is CE certified 7.

Total costs for clinical pilot study should not exceed €3000

8.

Data analysis is attainable with supplied software and/or is available with a University license

Based on the accelerometers we encountered in a broader literature search, eighteen devices were deemed suitable for the requirements analysis.

Alphabetically Supplementary Table S2 .

We found three devices to be eligible for use in a clinical setting (ActiGraph, AX3, and MOX1), which met all the technical requirements. The other fifteen devices did not meet the technical requirements. Decisive factors on device selection were based on sample rate with battery life and price. MOX1 (€195) had a battery life less than seven days with a sample rate between 60 and 100 Hz. The ActiGraph (€225) was almost twice as expensive as AX3 (€123). The AX3 was selected for a clinical pilot study, based on the combination of price, access to data, and no need for licensing or purchasing dedicated software.

A triaxial accelerometer measures the sum of inertial and gravitational acceleration in three directions [23] . It consists of a mass, suspended by a spring in a housing. Tri-axial accelerometers have three single axis accelerometers mounted in a box with sensitive axes in different directions, or as used in the AX3, have a sensor constructed using a single mass [24] . Acceleration is directly proportional to an external force and thus reflects the intensity and frequency of human movement. Accelerometers provide inclination sensing in response to gravity with respect to reference planes when the accelerometers rotate. The resulting inclination data is used to classify body posture orientations. Therefore, the measured accelerometry data can identify postures and daily movements using a classifier. The results can be related to a patient's functional status [25, 26] .

The positions and orientations of body segments are stable over time during static activities (e.g., lying, sitting, and standing). Static activities can therefore be identified by body segment orientation with respect to gravitation [27] . Positions and orientations of body segments vary over time during dynamic activities. During dynamic activities (e.g., walking, cycling, and walking stairs), body segments move naturally in a cyclical manner. For classification of activities, fixed-threshold classification [28] [29] [30] , reference-patternbased classification [27] , pattern-recognition strategies with statistical algorithms [27] , conventional or fuzzy logic [31] , and artificial neural networks [32, 33] were investigated. Machine learning algorithms, including deep learning algorithms, demonstrated high classification accuracy for human physical activity from tri-axial accelerometer data [34] [35] [36] .

The leg and arm were chosen for placement of the sensors, informed by recurring accelerometer placement for human activity recognition [37] and the expected comfort restrictions. Wound location, and expected added value of measuring both the upper and lower body to differentiate lying, sitting, and standing, were also considered. To investigate placement on the arm and leg, and the resulting accelerometer data, a pilot was run with 5 non-clinical subjects wearing 8 sensors. This context-dependent decision process led to the choice of placement of 2 sensors, as seen in Figure 1 . measured accelerometry data can identify postures and daily movements using a classifier. The results can be related to a patient's functional status [25, 26] .

The positions and orientations of body segments are stable over time during static activities (e.g., lying, sitting, and standing). Static activities can therefore be identified by body segment orientation with respect to gravitation [27] . Positions and orientations of body segments vary over time during dynamic activities. During dynamic activities (e.g., walking, cycling, and walking stairs), body segments move naturally in a cyclical manner.

For classification of activities, fixed-threshold classification [28] [29] [30] , referencepattern-based classification [27] , pattern-recognition strategies with statistical algorithms [27] , conventional or fuzzy logic [31] , and artificial neural networks [32, 33] were investigated. Machine learning algorithms, including deep learning algorithms, demonstrated high classification accuracy for human physical activity from tri-axial accelerometer data [34] [35] [36] .

The leg and arm were chosen for placement of the sensors, informed by recurring accelerometer placement for human activity recognition [37] and the expected comfort restrictions. Wound location, and expected added value of measuring both the upper and lower body to differentiate lying, sitting, and standing, were also considered. To investigate placement on the arm and leg, and the resulting accelerometer data, a pilot was run with 5 non-clinical subjects wearing 8 sensors. This context-dependent decision process led to the choice of placement of 2 sensors, as seen in Figure 1 . The next step towards classification is choosing an appropriate type of algorithm in combination with the provided (preprocessing of) data. Body acceleration was obtained by passing the acceleration signal through a high-pass filter with a cut-off frequency of 0.8 Hz. The measured data were divided into subsequent segments of 256 samples (2.56 s) based on an average person's walking cadence. For the selected activities, it was possible to efficiently provide correctly labeled data by giving participants instructions to perform all 6 activities in an ordered calibration session. Based on the provision of a labeled data The next step towards classification is choosing an appropriate type of algorithm in combination with the provided (preprocessing of) data. Body acceleration was obtained by passing the acceleration signal through a high-pass filter with a cut-off frequency of 0.8 Hz. The measured data were divided into subsequent segments of 256 samples (2.56 s) based on an average person's walking cadence. For the selected activities, it was possible to efficiently provide correctly labeled data by giving participants instructions to perform all 6 activities in an ordered calibration session. Based on the provision of a labeled data set and re-purposing existing open-source code for human activity recognition [38] , an artificial neural network (ANN) was implemented.

The ANN was trained by generating a number of features from the raw data stream. Previously indicated time/frequency-domain and heuristic features were used to this end, with presumed a high discriminatory ability and low computational cost [37] . For the training process, the following features previously described by Bunkheila [38] were included before selection: averaged total (i.e., static and dynamic) acceleration, root mean square of dynamic acceleration, autocorrelation of acceleration (height main peak, height, Sensors 2021, 21, 1979 5 of 17 and position of the second peak), spectral peaks (height and position of the first six peaks), and spectral power (total power of five adjacent predefined frequency bands).

In an attempt to further improve recognition by building on Janidarmian et al. [37] , the following features were also included: the median of total acceleration, the standard deviation of dynamic acceleration, and the median absolute deviation and magnitude of both total and dynamic acceleration. For simplicity, only the suggested time-domain features from [37] were included. For both sensors, data from all three axes were used, leading to a total of 160 features. The script used a feed-forward network comprised of one hidden layer with eighteen hidden nodes.

A neighbourhood component analysis (NCA) was applied as a feature selection technique. Only features with a feature weight ≥ 0.1 were kept using MATLAB's feature selection NCA implementation (fsnca [39] ), where the default solver type was SGD with an accompanying maximum of 30 iterations. This resulted in a feature selection of 59 features (see Supplementary Table S3 ). The sample size was reduced fitting the class with the lowest number of samples and then randomly selected. Static activities were performed more often than dynamic activities, to a ratio of 3 to 1. The results will be discussed in Section 3.1.2.

A standard K-fold (K = 10) and Leave-One-Subject-Out (LOO) cross-validation was used to determine the performance on the trained data. Where the latter is a better representation to investigate inter-subject differences, the former is a standard cross-validation, leading to comparative results but is susceptible to intra-subject data contamination. The main purpose of the study is alignment towards clinical implementation. This validation was thus not intended to be performance-driven but instead as verification for objective measurement of mobilization.

The MOV E M E NTT study (Mobilization Observation and Validation with acceleroMetry at thoraxceNTrum Twente) is a prospective single-center cohort study, reported as per the STROBE recommendations for observational research [40] . Adult patients undergoing cardiac surgery (CABG, valve surgery and combined CABG and valve surgery) between 2 June and 31 July 2020 were recruited at Thoraxcentrum Twente, Medisch Spectrum Twente, a tertiary nonacademic teaching hospital in Enschede, the Netherlands. Patients did not make additional hospital visits, and therefore conformed to hospital measures during the coronavirus pandemic.

Patients with a Katz Index of Independence in Activities of Daily Living functioning (KATZ-ADL) score larger than 2 before surgery (i.e., the patient is not independent in daily life mobilization) [41] , patients with an ICU stay exceeding 72 h, patients with postoperative cerebrovascular accidents, and patients mentally incompetent (including postoperative delirium) were excluded from the study. The sample size was determined to be 20 patients for sufficient evaluation of the accelerometer. This is larger than the median of 10 (1-36) patients recently reported in a large survey on accelerometer pattern recognition studies [35] .

The study ended at discharge from the hospital and thus no follow-up was obtained. This study was exempted from the Medical Research Involving Human Subjects Act by the Medical Ethics Committee Twente (METC Twente: K20-14) and was approved on 31 March 2020 by the local institutional review board.

Patients were recruited on hospital admission one day before cardiac surgery. After signing informed consent, two AX3 accelerometers were attached lateroproximal on the right upper arm and anterodistal on the right upper leg ( Figure 1 ). Hereafter, standardized exercises were performed, labelling measurements for lying, sitting, standing, walking, cycling, and walking the stairs (see Supplementary File 1 for the protocol); then the sensors were removed.

After surgery and ICU discharge, patients were transferred to the cardio-thoracic surgery ward. The sensors were attached, fixed with Tegaderm™ patches, and the mea-surements started, continuing 24/7. Measurements ended after 7 days or at earlier hospital discharge.

All patients received physical therapy twice a day until the 3rd postoperative day and then once a day. A standardized program starting from ICU discharge included:

• Day 1: Breathing exercises, coughing techniques, control mobilization of the upper and lower extremities, and transfer from bed to chair with assistance; • Day 2: Exercises as on Day 1. Self-transfer from bed to chair with or without assistance, and ambulation with assistance for 20 m at the surgical ward; • Day 3: Exercises as on Day 2. Ambulation with an increase in distance (±15 m) and frequency (3 times), cycling for 5-10 min at 0-10 Watt, depending on the hemodynamic stability; • Day 4: Exercises as on Day 3. Walking the stairs (1 floor) with assistance, information on home mobilization, and an increase cycling duration (5-10 min) and power (5-10 Watt).

Each patient-specific exercise program was based on the evaluation findings, comorbidities, and patient goals. Patients were stimulated to mobilize outside these sessions by nurses and other caregivers.

The primary endpoint of this study was the daily duration of each activity performed between 7 a.m. and 11 p.m. in minutes. The secondary endpoints were the length of ICU stay (nights) and surgical ward stay (nights). Routine baseline characteristics were determined based on the body mass index (BMI, kg/m 2 ) and European System for Cardiac Operative Risk Evaluation II (EuroSCORE II) definitions [42] .

Continuous variables were tested for normality with visual inspection of histograms and skewness/kurtosis measures. A p-value of less than 0.05 was set as statistically significant. Variables were analyzed with unpaired t-tests or Mann-Whitney U tests for continuous variables and for categorical variables with a Fisher Exact test.

Accuracy is defined as the ratio (%) of correct predictions to the total input samples. Precision, or positive predictive value, is the ratio (%) of correct classified samples of all samples belonging to the respective class. Recall, or the true positive rate, is the ratio (%) of correct classifications of the samples that should have been identified in that class.

A prespecified gender subgroup analysis, as encouraged by the Institute of Medicine [43] , was performed. After inclusion of 20 patients, only 15% was female. As this was lower than anticipated, the ethical review board approved the amended protocol to include up to 10 women (approved date: 9 July 2020). A linear mixed model with maximum likelihood was used to determine differences in activity time over time (day-to-day minutes and percentages), and the differences between male and female patients. Results are reported as the mean ± SD, median (with the 25th to 75th percentiles), or mean (min-max). Frequencies are reported as n (%).

Out of 75 patients selected for eligibility, in total 44 patients were excluded based on inability to perform the labelling measurement (n = 18), patient transfer from ICU to surgical ward planned during weekend days (n = 9), no available sensors (n = 8), patients referred from other hospitals (n = 6), or patients unwilling to participate (n = 3). Thus, 31 patients were included for the preoperative labelling measurements. Two patients had an ICU stay longer than 72 h and were excluded from the analysis. The CONSORT study flow diagram [44] is shown in Supplementary Figure S1.

Sensors 2021, 21, 1979 7 of 17 In this study, 22 male and 7 female patients were included, with no differences in their baseline (Table 2) or periprocedural (Table 3 ) characteristics, except for patient length (p < 0.001) and body surface area (p = 0.003). The majority of patients received isolated CABG (52%), 31% received valve surgery, and 17% CABG with concomitant valve surgery. Almost 47% of the performed CABG were handled off-pump, thus without cardiopulmonary bypass (7 out of 15 patients). The median ICU stay was 1 (1 to 2) night with no difference between male (1 (1 to 2)) and female (2 (1 to 2)) patients (U = 49, p = 0.13). Median surgical ward stay was 5 (3 to 6) nights, without a significant difference between male (5 (3 to 6)), and female (5 (3 to 7)) patients (U = 70, p = 0.73).

Thoracic drains were removed after 44 (39 to 47) hours after surgery. No perioperative myocardial infarction, prolonged intubation, readmission to the ICU, cerebral vascular accident, kidney failure, or gastrointestinal ischemia were observed. One pulmonary infection was observed in the male subgroup (4.5%), one rethoracotomy in both groups (male 4.5%, female 14%, p = 0.43) and one deep sternal wound infection occurred in the male group (4.5%). No patients died during hospital stay. Values are expressed as n (%) or median (25th to 75th percentiles). CABG = Coronary artery bypass grafting; ICU = Intensive care unit.

Labelled data from 31 participants was used for neural network training. Patients were asked to subsequently stand, sit on a chair, and lie on a bench on their back, right, and left side. Afterwards, patients walked back and forth on the corridor for 10 m, cycled on an exercise bike, and took one flight of stairs up and down. Each posture and activity had a duration of 30 s. On average, 11 m of labelled data were used for each patient. Segments of measured data were manually labelled by a researcher and plotted with MATLAB (see Figure S1 in Supplementary File 1). With visual inspection of the plots, the data samples were labelled as one of the six activities. For training purposes, a 50% overlap was used. No overlap was used for the actual postoperative activity classification.

The algorithm showed a 98% overall accuracy, as determined by a 10-fold crossvalidation and 96% for LOO validation using 160 or 59 features. All activities exceeded 90% on recall and precision (Table 4 ). The overall classification accuracy of the neural network, and that for male and female patients, started and remained constant at 95 to 96% throughout the study (Supplementary Figure S2 ). Both the recall and precision of the neural network was at least 78% (lying in bed) and increased over 90% throughout the study for all individual activities (Supplementary Figure S3 ).

Physical activities were obtained from 29 patients for 4 (2 to 5) days. One patient (3.4%) reported inconveniences regarding placement of the accelerometers. Patients spent 41 (20 to 62) min less time in bed for each consecutive hospital day, as determined by the mixed-model analysis (p < 0.001). For an overview of the relative time spent for all activities per hospital day, see Figure 2 .

Data are median (min-max).

The overall classification accuracy of the neural network, and that f patients, started and remained constant at 95 to 96% throughout the stu Figure S2 ). Both the recall and precision of the neural network was at leas and increased over 90% throughout the study for all individual activit Figure S3 ).

Physical activities were obtained from 29 patients for 4 (2 to 5) days reported inconveniences regarding placement of the accelerometers. Pati 62) min less time in bed for each consecutive hospital day, as determined b analysis (p < 0.001). For an overview of the relative time spent for all activit see Figure 2 . For all individual activities, see Table 5 for the fixed effects analysis and Figure 3 for the absolute and relative time spent for each activity per postoperative day. Thus, the model can be described by Equation (1):

where Y j is the patient activity as a function of the jth day, β 0 is the mean estimate of Day 1 (intercept), β 1 the increase of male over female patients, β 2 the slope of the time component, and Time is the number of days after Day 1 as a function of the jth day. The relative percentages of each activity changed during hospital stay (p = 0.01 for cycling and p < 0.001 for all other activities). For both absolute activity duration and relative percentage, no significant differences between men and women were observed (Table 5) .

Non-sedentary activities, such as standing, walking, and walking the stairs, increased during hospital stay (all p < 0.004), but not cycling (p = 0.07). From Day 3, the time spent cycling was stable (Figure 3i) . Accelerometers are not able to measure an increase in resistance for exercise bikes. Therefore, no improvement could be measured for cycling during this study.

Anthropometric parameters, such as patient length, weight, and body surface area, were not in the original protocol, but patient length and body surface area were significantly different between male and female patients. Patient length and body surface area were added as a single covariate to the linear mixed models for all primary endpoints. Length was a significant covariate in the model for lying in bed with only time (days after ICU discharge), F (1, 23.892) = 6.089, p = 0.02. Body surface area was also significant in a model for lying in bed, F (1, 46.359) = 4.173, p = 0.047.

Some patient movements were classified as cycling or walking the stairs on Day 1. According to the protocol and patient files, none of these activities were reported on Day 1. Misclassifications of the algorithm might occur when patients transfer between activities. As these transfers are of short duration and is similar for all patients for all postoperative days, this should have no effect on our interpretation of the results. Non-sedentary activities, such as standing, walking, and walking the stairs, increased during hospital stay (all p < 0.004), but not cycling (p = 0.07). From Day 3, the time spent cycling was stable (Figure 3i) . Accelerometers are not able to measure an increase in resistance for exercise bikes. Therefore, no improvement could be measured 

The aim of this study was to select and evaluate accelerometers for objective qualification of in-hospital mobilization after cardiac surgery with six static and dynamic patient activities. This study was designed in a practical manner, to be relevant for healthcare professionals, researchers, and patients.

After the literature review and expert consultation, three sensors were deemed suitable for objective qualification of in-hospital mobilization after cardiac surgery. The final choice for the AX3 sensor was mainly based on price, battery life, openly accessible sensor data, and not purchasing additional software. Many stakeholders were involved, including rehabilitation and sensor researchers, physical therapists, a cardiac surgeon, and a patient. In our current study, only accelerometer data were considered; future studies might consider different types of data to further improve classification [35] .

The aforementioned approach to measure mobilization activities appeared feasible. There were no technical problems with the AX3 devices as all data was usable for the analysis, and the Tegaderm™ patches did not unbind prematurely from the skin. One patient experienced some inconvenience from wearing the sensors. This study combined two accelerometers placed at both the upper and lower extremity. To the best of our knowledge, no research has been conducted with this approach to objectify in-hospital mobilization after cardiac surgery. All six activities were performed by patients and were recorded with the accelerometer.

In a previous study, an increase in the metabolic equivalent of tasks (METs) ≥3 and step count were observed after cardiac surgery in 83 patients receiving low-intensity physical therapy [12] . Light activities (METs < 3) increased from 3 ± 6 min on the 1st day after ICU discharge to 22 ± 24 min at Day 4. In our study, both light activities (standing and walking), and activities with moderate intensity (cycling and walking the stairs) were more frequently performed and earlier during postoperative stay. Light activities in our study were performed for 1 (1 to 5) min on the 1st day and increased to 55 (41 to 100) min on the 4th day. The difference can be explained by having standard physical therapy for all patients and using more active mobilization activities in our setting.

Previous studies reported higher functional levels and faster mobilization for male patients [12, 13] . Men might have a more competitive attitude to mobilization or overestimate self-reported activities. Women also might be more hesitant to mobilize after surgery. In this study, no effect of sex was observed for any of the activities observed.

Some of the baseline and perioperative and postoperative variables might affect mobilization activities. The original study protocol did not include patient length and body surface area. It turned out, as expected, that these parameters are significantly different between male and female patients. Both factors were significant covariates in a linear mixed model for quantifying lying in bed only. Both parameters remained significant in a model with sex as fixed factor. However, lying in bed was still not significant between male and female patients (p = 0.08). Future studies should include anthropometric parameters, i.e., patient length, weight, BMI, and body surface area, to correct for differences in the baseline characteristics between male and female patients.

To generate hypotheses for future studies, all parameters from Tables 2 and 3 were added to the linear mixed model as a fixed factor or covariate. Although p-values < 0.05 can be considered as significant, with these multiple comparisons, type I errors are likely to occur. Only length and body surface area were significant covariates for the static activities. For the dynamic activities, only previous neurological dysfunction had an impact on two activities: walking (both absolute and relative) and walking the stairs (both absolute and relative). The results from the mixed models are added as Supplementary File 2.

An accelerometer for measuring inpatient mobilization was tested for 29 patients. A recent survey showed that previous studies included a median of 10 (1-36) patients [35] . The neural network precision and recall per activity for this study exceeded 78%, as determined with LOO validation over n number patients, even from n = 2. After n = 10, both the precision and recall for all individual activities exceeded 90%, and remained stable (Supplementary Figure S3) . Nonetheless, the difference between K-fold (K = 10) and LOO cross-validation, especially looking at the range and difference in training size, might warrant careful inspection. This calls for future work to investigate an analysis on unseen subjects. The sample size with the calibration step included is deemed sufficient to draw conclusions on the neural network performance. No differences in neural network performance between male and female patients were observed.

The study was designed for measuring, with two accelerometers, all six activities after cardiac surgery. As a sensitivity analysis, the accuracy of only one accelerometer was also calculated. With the upper arm sensor only, an accuracy of 80% (77-83) was obtained with K-fold cross validation, and 74% (17-95) with LOO validation. For the upper leg, the neural network scored the activities accurate for 95% (94-95) and 92% (76-100), respectively. For future studies, only one accelerometer on the upper leg should be considered for measuring the postoperative mobilization activities after cardiac surgery. It is likely that our approach is suitable for classification of mobilization for other types of surgery or other patient groups, as long as sensor placement does not affect the wound area. Furthermore, these in-hospital sensor data might prove beneficial as a baseline measurement for the intake of outpatients for cardiac rehabilitation.

Preoperative labeling might have introduced misclassifications as all patients walked without support. Patients in their early postoperative stay walk slower and sometimes with a walker or physical therapist support. This might have overestimated standing and underestimated walking. Postoperative activity forms were designed to estimate mobilization activities from a professional perspective, and compare these with the classifications. These forms were too generic to detect misclassifications. However, the data retrieved from the forms were consistent with the actual classifications.

For time-efficiency reasons, only the reduced feature set was used to classify the activities after surgery, although in retrospect a more thorough NCA threshold choice for feature selection is advised for future work [45] . The cutoff ≥0.1 was deemed appropriate based on MATLAB's fcnca function [39] , and visual inspection of the histograms. No sensitivity analysis was performed on setting the cutoff for features larger than zero, but which is recommended for future research.

Patients were discharged from ICU staring at 10:30 a.m. and therefore the 1st day measurements missed some hours in the absolute data analysis. Indeed, low absolute time spent on lying in bed and sitting (Figure 3a ,c) was observed on Day 1. Using proportional time (Figure 3b ,d) might reflect the activities better on days with incomplete data collection. Using both parameters should be used for future studies as well.

Small positional changes of the sensors have strong effects on the sensor output [46] . Before our clinical pilot study started, four positions were considered on both the arm and leg. The best positions were identified in a calibration study in four healthy subjects. Potential sensor position changes during the study were not identified. To increase reliability, a user-friendly position of the sensor, no clipping of true accelerations, and a tight body fit were recommended [46] . The sensor placement in our study on the upper leg might warrant further inspection from a biomechanical standpoint due to possible inclusion of soft tissue (e.g., muscle movement). Importantly, this sensor in combination with placement on the upper arm did not hamper free movement. Furthermore, 97% of the included patients did not report inconvenience. Sensors for the preoperative calibration measurements and postoperative quantification were attached by one researcher according to prespecified landmarks. With this approach, a tight fit and similar positions were achieved.

The clinical pilot was designed to classify in-hospital mobilization activities. Future research should investigate how this information can be used to improve rehabilitation. Physical therapists and nurses can tailor their strategies with reliable information on performed activities. With persuasive designs and serious gaming, patients could be motivated to achieve higher levels of functional independence. New technical and clinical requirements might be formulated, because direct feedback is different from analysis after hospital discharge. The next step is to provide feedback to patients and healthcare professionals, to speed up recovery.

Commercially available devices for in-hospital mobilization after cardiac surgery were screened for suitability in a literature review and one device selected for evaluation. Static and dynamic activities, e.g., lying, sitting, standing, walking, cycling, and walking the stairs, were classified with an AX3 accelerometer for postoperative patients at a cardiac surgery ward. Patients spent 41 (20 to 62) min less time in bed for each consecutive hospital day (p < 0.001) based on the sensor data. Non-sedentary activities, such as standing, walking, and walking the stairs, increased during the hospital stay (all p < 0.004), but not cycling (p = 0.07). No differences between male and female patients were observed for all endpoints.

The approach presented in this study deems applicable for monitoring postoperative recovery of patients in similar contexts.

Supplementary Materials: The following are available online at https://www.mdpi.com/1424-8 220/21/6/1979/s1, Figure S1 : CONSORT patient flow diagram of MOV E M E NTT study, Figure S2 : Overall neural network accuracy using n number of patients in training data set, Figure S3 : Neural network recall and precision per activity using n number of patients in training data set, Table S1 : Technical and clinical requirements for devices used to classify inpatient physical activity after cardiac surgery, Table S2 : Application of technical and clinical requirements on 18 wearable devices, Table S3 : Feature selection after neighbourhood component analysis, File S1: Neural network training protocol, File S2: Results from mixed model analysis.

The Society of Thoracic Surgeons Adult Cardiac Surgery Database: 2019 Update on Outcomes and Quality

Perspectives of patients and health-care professionals on physical activity of hospitalized patients

I'm not going to walk, just for the sake of walking

Functional impact of 10 days of bed rest in healthy older adults

Exploring barriers to physical activity of patients at the internal medicine and surgical wards: A retrospective analysis of continuously collected data

Physical activity during hospitalization: Activities and preferences of adults versus older adults

Exercise-Based Cardiac Rehabilitation for Coronary Heart Disease: Cochrane Systematic Review and Meta-Analysis

Cardiac Rehabilitation Utilization During an Acute Cardiac Hospitalization: A NATIONAL SAMPLE

Physiotherapy mobility and walking management of uncomplicated coronary artery bypass graft (CABG) surgery patients: A survey of clinicians' perspectives in Australia and New Zealand

Effects of early mobilisation in patients after cardiac surgery: A systematic review

The effect of early mobilization protocols on postoperative outcomes following abdominal and thoracic surgery: A systematic review

Effect of physical activity in the first five days after cardiac surgery

Postoperative mobilization after total hip arthroplasty: Measured by Fitbit activity trackers. Der Orthop

Decreased physical activity is a predictor for a complicated recovery post cardiac surgery. Health Qual

Preliminary evidence for physical activity following pelvic exenteration: A pilot longitudinal cohort study

Construct validity and inter-rater reliability of the Dutch activity measure for post-acute care "6-clicks" basic mobility form to assess the mobility of hospitalized patients

Supervised moderate intensity exercise improves distance walked at hospital discharge following coronary artery bypass graft surgery-A randomised controlled trial

Promoting mobility and reducing length of stay in hospitalized general medicine patients: A quality-improvement project

Assessing pre-and postoperative activity levels with an accelerometer: A proof of concept study

Exercise therapy after coronary artery bypass graft surgery: A randomized comparison of a high and low frequency exercise therapy program

Measuring physical activity levels in hospitalized patients: A comparison between behavioural mapping and data from an accelerometer

Functional recovery in the elderly after major surgery: Assessment of mobility recovery using wireless technology

Inertial Sensing of Human Movement

Design, realization and characterization of a symmetrical triaxial capacitive accelerometer for medical applications

A review of accelerometry-based wearable motion detectors for physical activity monitoring

Classification of basic daily movements using a triaxial accelerometer

Detection of static and dynamic activities using uniaxial accelerometers

Physical activity monitoring based on accelerometry: Validation and comparison with video observation

Assessment of posture and motion by multichannel piezoresistive accelerometer recordings

Motion pattern and posture: Correctly assessed by calibrated accelerometers

Wearable sensors and telerehabilitation

Incline, speed, and distance assessment during unconstrained walking

Computerized analysis of daily life motor activity for ambulatory monitoring

Machine learning methods for classifying human physical activity from on-body accelerometers

Deep learning for sensor-based activity recognition: A survey

Deep, Convolutional, and Recurrent Models for Human Activity Recognition using Wearables

A comprehensive analysis on wearable acceleration sensors in human activity recognition

Signal Processing and Machine Learning Techniques for Sensor Data Analytics". MATLAB Central File Exchange

Feature Selection Using Neighborhood Component Analysis for Classification-MATLAB Fscnca

Strengthening the Reporting of Observational Studies in Epidemiology (STROBE): Explanation and elaboration

Studies of Illness in the Aged. The Index of Adl: A Standardized Measure of Biological and Psychosocial Function

Board on Population Health and Public Health Practice

Updated guidelines for reporting parallel group randomised trials

Smartphone Based Human Activity Recognition with Feature Selection and Dense Neural Network

Optimal sensor placement for measuring physical activity with a 3D accelerometer

The authors acknowledge all physical therapists, nurse practitioners and nurses from the cardio-thoracic surgery ward for their study activities. The human body in the graphical abstract was designed by macrovector/Freepik.

The authors declare no conflict of interest. The funder had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or in the decision to publish the results.Sensors 2021, 21, 1979