Wearables accurately assess dependence in older adults

With the help of various machine learning (ML) algorithms, wearable devices can semi-automatically acquire and analyse data for the assessment of dependence in older adults, according to a recent study.

“The use of wearables and ML contribute to create a holistic and semi-ecological model that improves the traditional assessment of instrumental activity of daily living (IADL) performance,” the researchers said. Such an approach could not only help reduce the workload but also help minimize healthcare costs.

The study included 78 older adults (69 women), who were asked to complete a shopping activity task while wearing two mobile sensor devices in the dominant hand. The shopping task consists of 14 shopping stages throughout which a trial supervisor collected accelerometer, heart rate, electrodermal, and gyroscope data through a smartphone app via Bluetooth. IADL was assessed using the Lawton and Brody Scale.

After processing the data using three ML algorithms (k-Nearest Neighbours [NN], Random Forest, and Support Vector Machines [SVM]), the researchers found that using wearables could indeed replace traditional questionnaires for data collection for the assessment of IADL dependence. In particular, when making use of at least 10 sensor features, all algorithms achieved an F1 score >90 percent, indicative of good overall accuracy. [Int J Med Inform 2022;157:104625]

The simple 1-NN algorithm, using only 11 features, emerged as the best predictive model, yielding an F1 score of 1.00. Accuracy, sensitivity, and specificity estimates were likewise 1.00, suggesting potentially perfect performance.

Notably, a lighter 1-NN model using only five sensor features also showed excellent overall performance, with an F1 score of 0.963 and an accuracy estimate of 97.16 percent. Corresponding sensitivity and specificity values for this model were 97.39 percent and 97.03 percent, respectively.

Moreover, an SVM model using 65 features yielded good overall performance, with an F1 score of 0.98, accuracy of 96.63 percent, sensitivity of 97.85 percent, and specificity of 99.11 percent. The RF algorithm likewise achieved good performance when using 69 total features, producing an F1 score of 0.95, accuracy of 96.61 percent, sensitivity of 92.67 percent, and specificity of 98.96 percent.

“For the sample analysed in this paper—keeping sensitivity greater than specificity—we created an automatic model with data from only two wearables and an accuracy over 99 percent, and another model with data from only one wearable and an accuracy of 97 percent,” the researchers said. “Therefore, it is possible to substitute/replace the manual questionnaires by the automatic assessment of dependency proposed in this paper with high accuracy.”

Future efforts should focus on making the current system completely ecological and automated, doing away with the need for an external observer, they added. This would require that the annotation of shopping stages also be automated, though the researchers said that such a move is possible.

In turn, complete automation “would further reduce economic and human costs in public and private health systems,” the researchers said. Such an automated approach could be integrated into a generalized m-health system to allow interoperability across different devices, applications, and sensors.