Current State of Non-wearable Sensor Technologies for Monitoring Activity Patterns to Detect Symptoms of Mild Cognitive Impairment to Alzheimer’s Disease

Table 2

General characteristics of the studies included in group 2.

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
Lussier et al. [28]	(i) Cross-sectional study; (ii) Examine how strongly the sensor-based activity data was related to the MCI clinical diagnosis, cognitive performance, and performance-based measures	Participants: CH: 26; MCI: 22; mean age: NS; duration: NS	bADL, iADL	Regression analysis	Sensor readings were used to calculate the duration of the scripted tasks performed by the participants. Single trial at two different sites was reported, and there was no relevance of a time series treatment of measures derived from sensor data.	NS	Intersite validation of sensor data was examined through ANOVA: both sites were comparable in terms of time spent in each living area as well as for the use of domestic appliances and storage.	Sensor-based activity data were associated with memory and executive performances and significantly contributed to the prediction of MCI.
Arifoglu et al. [29]	(i) Longitudinal and cross-sectional study; (ii) Recognize activity patterns from sensor-based activity data and detect abnormal behavior related to dementia	Longitudinal study participants: CH: 1; cross-sectional study participants: CH: 20; mean age: NS; duration: not clear	bADL, iADL, sleep	CNN, LSTM	Sensor data is discretized by sliding window approach with a constant time-slice length of 60 sec and forming a time series chunk of tXf matrix where rows are time slices and columns are sensor readings. This preprocessed data was given as input for CNN for classification.	NS	Cohen’s Kappa statistics was computed in order to show the robustness of the proposed CNN-2D classifier. A value of 0.64431 indicated a substantial agreement.	When both temporal and feature dimensions were considered, results of activity recognition and anomalous pattern detection were found to be promising, especially in case of repetition-related activities and confusion-related activities.
Paudel et al. [30]	(i) Longitudinal study; (ii) From the sensor-based activity data, derive behavioral/activity features and determine any anomalous activity patters that could be predictive of onset of dementia	Participants: CH: 5; MCI: 5; age: 80-91 years; duration: not clear	bADL, iADL, sleep	Logistic regression, LDA, DT, SVM, KNN, RF, Ada Boosting, one-class SVM	From the raw sensor data stream, features related to each activity were derived and that would serve as data points for modeling. These data points were treated as discrete data points rather than any time series while training and testing the models.	NS	NS	Regardless of any specific activity domain, features derived from all activities contributed significantly to classify the subject into healthy or MCI group.
Gochoo et al. [31]	(i) Longitudinal study; (ii) From sensor-based activity data, derive Martino-Saltzman’s (MS) travel patterns that could be indicator of individual’s cognitive state	Participants: CH: 1; mean age: NS; duration: 625 days	Movement patterns	Deep CNN, NB, GB, RF	Raw dataset representing a long list of consecutive movements (motion sensor data) is segmented into groups of travel episodes. Episode starts when there is any movement is occurred in the raw sensor data after the end of previous episode and the episode stops if there is no motion for more than 10 seconds. Each episode was converted to a binary image which was input to the DCNN model.	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	Each episode was classified into one of the patterns—direct, pacing, lapping, and random. These results could be precursor to further clinical evaluation and diagnosis.
Li et al. [32]	(i) Cross-sectional study; (ii) Perform a quick screening of older adults and classify into dementia or nondementia group utilizing motion trajectory-based features derived from iADL data	Participants: CH: 72; MCI: 43; dementia: 7; age: 60–74 years; duration: NS	bADL, iADL	Bayesian network, SVM, RF, NB	As the participants performed their activities, the IDs of motion sensors triggered for each activity were combined together to form a motion “Trajectory” of that activity. Appropriate features were extracted from these trajectories to be fed into learning algorithm.	NS	NS	Wandering patterns such as pacing and lapping (as represented trajectory features) were significantly different between subjects with dementia and without dementia. Subjects classified into dementia group could not be differentiated between MCI and CH perhaps due to nature of activities performed in very short span of time.
Dawadi et al. [33]	(i) Cross-sectional study; (ii) Automatically calculate the activity performance score from sensor captured activity data; correlate this score with expert assigned as well as predict cognitive health condition based on this automated score	Participants: CH: 145; MCI: 32; Dem: 2; mean age: not clear; multiple age groups- yo; 45-59 yo; 60-74 yo; yo; duration: 1 hour	bADL, iADL	SVM (bagged), NN, NB	From sensor event data stream, the activities (scripted tasks) were recognized, and subsequently, activity features were computed which would indicate the quality of task completion and quantum of parallelism. Single trial at a single site was reported, and there was no relevance of a time series treatment of measures derived from sensor data.	NS	Two trained neuropsychologists observed the participants performing the tasks and recorded two scores, namely, task accuracy and sequencing score. The sensor-derived task features were examined to be correlated to these observers rated scores. Interrater reliability agreement came out to be 97.88% and 99.57% for the accuracy and sequencing scores, respectively.	The correlation () between smart home sensor-derived features and task accuracy scores was found to be statistically significant (rather than task sequencing score). While predicting cognitive health, study was able to classify between CH and dementia with a better accuracy than classifying between CH and MCI.

ADL: activities of daily living; a-MCI: amnestic MCI; bADL: basic ADL; CH: cognitively healthy; CNN: Convolutional Neural Network; DT: decision tree; GB: gradient boost; HMM: Hidden Markov Model; iADL: instrumental ADL; KNN: K nearest neighbors; LDA: linear discriminant analysis; LSTM: long short-term memory; MCI: mild cognitive impairment; na-MCI: nonamnestic MCI; NB: Naïve Bayes; NN: neural network; NS: not specified; RF: Random Forest; RNN: recurrent neural network; SVM: Support Vector Machine.