Adherence Monitor for Measurement of Removable Cast Walker Wear-Time: Multiple Sensors and Predictive Analytics Improve Accuracy

Abstract

Background: Treatment of diabetes and its complications is a primary health care expense. Up to 25% of people with diabetes will develop diabetic foot ulcers (DFUs). Removable cast walker (RCW) boots commonly prescribed for DFU treatment, promote healing, and provide offloading and wound protection. Patient RCW removal for hygiene and wound care can lead to decreased adherence and treatment effectiveness. This study evaluated a new system for wear-time adherence measurement using multiple sensor types.

Methods: An electronic wear-time monitor was developed, which included internal and external temperature sensors, an accelerometer, and capacitive proximity foot and ankle sensors. Time-stamped and date-stamped data were saved once per minute for up to 22 days. Ten healthy volunteer subjects were recruited to wear an RCW for two weeks while keeping a diary of don/doff times. Sensor data were then compared with volunteers' wear diaries using confusion matrix predictive analytics.

Results: Algorithms were developed for data processing. Correlation coefficients between algorithms and diaries were calculated for individual and multiple sensor combinations. Differential temperature and accelerometer algorithms were significantly better at predicting subject wear-time than individual temperature sensor algorithms (P = .009, P = .001, respectively). Foot proximity had significantly better correlation with subject diaries than temperature (P = .024), and acceleration algorithms (P = .005). Multi-sensor analysis showed high correlation (.96) with wear-time from subject diaries.

Conclusions: Removable cast walker wear-time can be accurately determined using an electronic data recorder and multiple sensors. Wear-time measurement accuracy can be improved using algorithms that operate on data from multiple sensors that use a variety of sensor technologies.

Keywords: adherence; adherence monitor; diabetic foot ulcer; orthosis wear-time; removable cast walker; wear-time.

Figure 1.

(a) Removable cast walker (Aircast AirSelect, DJO Global, Vista, California) with the adherence monitor mounted externally over the ankle. (b) Inside of the RCW with liner removed and right-side air bladder lifted out of the way to show sensor placement. (P) shows placement of proximity sensors. (T) shows placement of internal temperature sensor.

Figure 2.

Mean Matthews correlation coefficient (MCC) results with 95% confidence intervals for 13 subject trials. The blue dashed line is drawn at 0.87 (87%), the level of predictive performance provided by the combination of the best acceleration and the best temperature algorithm (ie, acceleration 4 and temperature 2). (a) Algorithms which produced the lowest accuracy. (b) Best performing sensor algorithms. (c) Exclusive combinations of sensor/algorithms at predicting subject wear-time. (d) Non-exclusive combinations, or majority agreement of sensor/algorithms at predicting subject wear-time.

Figure 3.

Examples of inaccurate wear/non-wear classification using differential temperature algorithms. The top graph shows internal and ambient temperature of three thermal events. The bottom graph shows data from the subject’s diary and the output of the two temperature algorithms (high = don, low = doff). The first two temperature events which were erroneously labeled as don time by the temperature algorithms, occurred during transportation of the RCW in a warm vehicle. The third event likely occurred because of loosely wearing the RCW in an office setting, such that, the leg was not in contact with the boot liner and hence was not near the internal temperature sensor.

Figure 4.

Examples of inaccurate wear-time classification by an accelerometer algorithm. Vehicle transportation of the RCW or other non-wear activities was erroneously interpreted as wear-time by the best accelerometer wear-time algorithm. The top graph shows accelerometer X, Y, Z data. The subject diary trace shows actual don/doff times. The accelerometer algorithm trace shows premature and incorrect identification of the start and end of wear event 1, and good identification of event 3. From the subject’s diary, events 2 and 4 are known to be transportation of the orthosis but were erroneously labeled as don time by the algorithm.