Uncovering Social States in Healthy and Clinical Populations Using Digital Phenotyping and Hidden Markov Models: Observational Study

Abstract

Background: Brain-related disorders are characterized by observable behavioral symptoms, for example, social withdrawal. Smartphones can passively collect behavioral data reflecting digital activities such as communication app usage and calls. These data are collected objectively in real time, avoiding recall bias, and may, therefore, be a useful tool for measuring behaviors related to social functioning. Despite promising clinical utility, analyzing smartphone data is challenging as datasets often include a range of temporal features prone to missingness.

Objective: Hidden Markov models (HMMs) provide interpretable, lower-dimensional temporal representations of data, allowing for missingness. This study aimed to investigate the HMM as a method for modeling smartphone time series data.

Methods: We applied an HMM to an aggregate dataset of smartphone measures designed to assess phone-related social functioning in healthy controls (HCs) and participants with schizophrenia, Alzheimer disease (AD), and memory complaints. We trained the HMM on a subset of HCs (91/348, 26.1%) and selected a model with socially active and inactive states. Then, we generated hidden state sequences per participant and calculated their "total dwell time," that is, the percentage of time spent in the socially active state. Linear regression models were used to compare the total dwell time to social and clinical measures in a subset of participants with available measures, and logistic regression was used to compare total dwell times between diagnostic groups and HCs. We primarily reported results from a 2-state HMM but also verified results in HMMs with more hidden states and trained on the whole participant dataset.

Results: We identified lower total dwell times in participants with AD (26/257, 10.1%) versus withheld HCs (156/257, 60.7%; odds ratio 0.95, 95% CI 0.92-0.97; false discovery rate [FDR]-corrected P<.001), as well as in participants with memory complaints (57/257, 22.2%; odds ratio 0.97, 95% CI 0.96-0.99; FDR-corrected P=.004). The result in the AD group was very robust across HMM variations, whereas the result in the memory complaints group was less robust. We also observed an interaction between the AD group and total dwell time when predicting social functioning (FDR-corrected P=.02). No significant relationships regarding total dwell time were identified for participants with schizophrenia (18/257, 7%; P>.99).

Conclusions: We found the HMM to be a practical, interpretable method for digital phenotyping analysis, providing an objective phenotype that is a possible indicator of social functioning.

Keywords: Alzheimer disease; cognitive impairment; digital phenotyping; hidden Markov model; mHealth; mobile health; mobile phone; passive monitoring; schizophrenia; smartphone; social behavior.

Figure 1

Overview of the hidden Markov model (HMM) approach showing the main processing and modeling steps involved in the method. (A) The Behapp app was installed and collected data passively. (B) These data were processed into activity bins. (C) The HMM was trained on the binned hourly time series. (D) The hidden state sequence was generated for each validation participant and their total dwell time calculated. (E) The total dwell time was compared to clinical measures. (F) Lower socially active dwell time in AD versus HCs, and an interaction between socially active dwell time and AD when predicting social functioning, were observed. AD: Alzheimer disease; HC: healthy control; SCC: subjective cognitive complaints; SZ: schizophrenia; S1: state 1; S2: state 2; z_t: hidden state at time point, t.

Figure 2

Age density distributions for validation participants. (A) Distribution of ages for all validation participants. (B) Distribution of ages for validation participants with social measures. Plotted using kernel density estimation.

Figure 3

Emission probabilities of the selected 2-state model. Emission probabilities are provided for (A) state 1 (S1) and (B) state 2 (S2).

Figure 4

Examples of which behaviors may correspond to the hidden states. For the socially active state, various social behaviors are displayed, including calls and app use; in the socially inactive state, there may be no phone usage or phone usage without corresponding social behaviors.

Figure 5

Example time series with high social activity. The observed time series composed of hourly bins (bottom 5 rows) of a participant compared with their corresponding predicted hidden state sequence (top row). S1: state 1 (socially inactive state); S2: state 2 (socially active state).

Figure 6

Example time series with low social activity. The observed time series composed of hourly bins (bottom 5 rows) of another participant compared with their corresponding predicted hidden state sequence (top row). S1: state 1 (socially inactive state); S2: state 2 (socially active state).

Figure 7

An example of a 2-day period of a participant’s time series. This participant showed higher social activity during the daytime than the nighttime. 0: midnight; S1: state 1 (socially inactive state); S2: state 2 (socially active state).

Figure 8

The probability of transitioning into the socially active state from each state, for each hour in the day. 0: midnight; S1: state 1 (socially inactive state), S2: state 2 (socially active state).

Figure 9

Social functioning scale score against total dwell time, with interactions displayed for the different groups.

Figure 10

A box plot of the total dwell times per participant for the different diagnostic groups. There is a significant difference between the HC and AD groups, and the HC and SCC groups. AD: Alzheimer disease; HC: healthy control; SCC: subjective cognitive complaints; SZ: schizophrenia.