|
| Study | Study design; main objective | Participants; mean age; mean data collection duration | Activities in focus | Machine learning technique | Nature of data gathered/preprocessed | Artifacts detection/correction in sensor data gathered | Approach to address variability, reliability, etc. | Main results |
|
| Alberdi et al. [22] | (i) Longitudinal study
(ii) Analyze the correlation between activity features and cognitive/behavioral health assessment scores
(iii) Detect cognitive/behavioral symptoms that could be indicator of AD | Participants: CH: 13; at risk: 10; MCI: 6; mean age: 84.35; duration: 19.95 months | bADL, iADL, sleep, other daily routines | Regression: SVR, LR, KNN; classification: SVM, AdaBoost, MLP, RF | From sensor event data stream, the activities were recognized, and subsequently activity/behavioral features were computed at day level, and they formed a “time series” data. From this “time series” data, summary statistical features were computed using a sliding window approach for further analysis. | Gaussian detrending—to remove the effect of nonstationary components (e.g., periodic components) in the time series | Reliable change index (RCI) (standardization method) was computed to address intersubject variability in health assessment scores. RCI computation of the test scores in this study utilized test-retest reliability and standard deviation values that the tests have shown in their development cohorts and/or in previous work. | Sensor-based activity observations such as sleep and overnight patterns along with daily routine features contributed significantly to the prediction of various cognitive assessment scores. |
| Schinle et al. [23] | (i) Longitudinal study
(ii) Detect possible indicators of onset of dementia using individuals’ day-night rhythm and night-time activity as relevant parameters | Participants: 10; mean age: NS; duration: 13 months | Movement patterns to infer sleep patterns | Local outlier factor (clustering) | From the raw sensor events data stream, two types of events were derived, namely, motion events and outside home events. Based on density of these events, three measures such as wake-up time, bed time, and night time activity count were determined. Each of these measures formed a time series data for learning behavioral profile and trend. | Nothing specific was discussed in terms of artifact detection and correction. However, authors noted that the single motion sensor could not detect many activities outside of sensor’s vision (e.g., latent motion) and that would lead to the understanding of “no movements” impacting the accuracy of prediction. | NS | From the activity trend data, the wake-up times/bed times were recognized, classified as no anomaly or slight anomaly or severe anomaly, based on anomaly detection rules defined. These rules were defined heuristically after examining the distribution of the wake-up and bed times for several households. |
| Sharma et al. [24] | (i) Longitudinal study
(ii) Detect early symptoms of MCI using ADL data from sensors
(iii) Address to fill in data gaps caused by sensor failures | Participants: 50; mean age: NS; duration: 6 months | Daily routine pattern | RNN | Expressed human routine as a time series inference based on the raw data stream from sensors. | Data gaps (missing sensor values) caused due to faulty sensors/dead sensors were filled in using time series prediction techniques such as RNN. Missing sensor values indicate the presence of certain noise in the incoming sensor data stream. | NS | All the activities recognized from sensor data on a particular day of a participant were compared against the same participant’s data from previous day to compute the deviation; if this deviation was noted for more than month, abnormality was detected and referred for further clinical evaluation. |
| Akl et al. [25] | (i) Longitudinal study
(ii) Home-based automatic detection of MCI symptoms through individual’s room activity distributions | Participants: CH-59; a-MCI: 11; na-MCI: 15; mean age: not clear; duration: 3 years average | General home activity patterns | Affinity propagation (clustering) | Activity data from raw sensor data stream was used to compute room activity probability distributions. These probability distributions were considered as discretized values over a fixed time interval but not treated as a time series data. | Discarded sensor readings of certain days when unusual activity patterns (zero, too many, etc.) were seen and those could be due to variety of reasons such as study participant had visitors or some sensor failed, etc. | NS | Individual’s activity distributions, combined from all the four rooms, namely, bedroom, bathroom, living room, and kitchen were found to be significant contributor in predicting an individual as “Cognitively intact” or “MCI.” Within MCI class, individuals transitioned to subtype “a-MCI” showed significant changes in their bedroom activity distributions that were mainly attributed to disturbed sleep patterns. |
| Akl et al. [26] | (i) Longitudinal study
(ii) Autonomous detection of MCI symptoms based on walking speed and general activity recognized through unobtrusive sensing technology | Participants: CH: 79; MCI: 18; mean age: not clear (70 and above); duration: 171.9 weeks | bADL, other daily routines | SVM, RF | From the raw sensor events data stream, walking-related predefined measures were computed, and these measures were transformed into features using signal processing approach (based on sliding window). These features served as data points for machine learning modeling and prediction. | Discarded sensor readings of certain days when unusual activity patterns (zero, too many, etc.) were seen and those could be due to variety of reasons such as study participant had visitors or some sensor failed, etc. | NS | Trajectories of weekly walking speed-related measures and the participant’s age and gender were the most important for detecting MCI in older adults. The feature, “trajectories of measures,” refers to the concatenation of corresponding measures as they appeared in each window (“l” week). |
| Albeiruti et al. [27] | (i) Longitudinal study
(ii) Detect sudden and gradual abnormalities in behavior of older adults based on movement pattern recognized through simple motion sensors | Participants: CH: 1; mean age: NS; duration: 219 days | Movement patterns | HMM | Sensor firing states (from raw sensor data) were treated as observed states in learning the Hidden Markov model’s parameters which characterized the subject’s behavior. Probability distributions of a sensor firing after another sensor were derived from observed data (training data). | A manual cleanup was done to discard the sensor readings where an “ON” state was observed without having a corresponding “OFF” state. | NS | The behavior model (movement pattern based) was able to detect abnormalities on specific days within the monitoring period. Abnormal days could be indicator of cognitive decline. |
|
