Use of passively collected actigraphy data to detect individual depressive symptoms in a... 2024 Price et al

[Andy]

Full title: Use of passively collected actigraphy data to detect individual depressive symptoms in a clinical subpopulation and a general population.

The presentation of major depressive disorder (MDD) can vary widely due to its heterogeneity, including inter- and intraindividual symptom variability, making MDD difficult to diagnose with standard measures in clinical settings. Prior work has demonstrated that passively collected actigraphy can be used to detect MDD at a disorder level; however, given the heterogeneous nature of MDD, comprising multiple distinct symptoms, it is important to measure the degree to which various MDD symptoms may be captured by such passive data.

The current study investigated whether individual depressive symptoms could be detected from passively collected actigraphy data in a (a) clinical subpopulation (i.e., moderate depressive symptoms or greater) and (b) general population. Using data from the National Health and Nutrition Examination Survey, a large nationally representative sample (N = 8,378), we employed a convolutional neural network to determine which depressive symptoms in each population could be detected by wrist-worn, minute-level actigraphy data.

Findings indicated a small-moderate correspondence between the predictions and observed outcomes for mood, psychomotor, and suicide items (area under the receiver operating characteristic curve [AUCs] = 0.58–0.61); a moderate-large correspondence for anhedonia (AUC = 0.64); and a large correspondence for fatigue (AUC = 0.74) in the clinical subpopulation (n = 766); and a small-moderate correspondence for sleep, appetite, psychomotor, and suicide items (AUCs = 0.56–0.60) in the general population (n = 8,378).

Thus, individual depressive symptoms can be detected in individuals who likely meet the criteria for MDD, suggesting that wrist-worn actigraphy may be suitable for passively assessing these symptoms, providing important clinical implications for the diagnosis and treatment of MDD.

Impact Statement
The coupling of deep learning methods with passive monitoring of an individual’s naturalistic movement provides a unique opportunity to detect depressive symptoms without the necessity for frequent clinical visits or self-report measures. The present work builds upon previous efforts to evaluate which depressive symptoms are best captured by passively collected physical activity data, and how this differs between individuals in the general population and individuals who meet criteria for depression.

Our findings provide insight into which individual depressive symptoms may be best detected by passively collected physical activity data, providing important assessment and treatment implications for depression.

Paywall, https://psycnet.apa.org/doiLanding?doi=10.1037/abn0000933

 

[Haveyoutriedyoga]

I think you have accidentally posted plans for a new episode of Black Mirror..

 

[Hutan]

Could be useful to have a look to see what they measured with respect to fatigue. The authors seem quite positive about the capacity for measures to predict fatigue.

 

[Haveyoutriedyoga]

If this is the pre-print then it looks as though they simply got patients to do the PHQ-9 which has a question on how much you feel tired or have little energy, and compared that to the actigraphy data.

 

[rvallee]

Aside from being a terrible idea with no possible good outcome, studies like this just shows how medicine is as totally clueless as ever about depression and has not actually made any progress in understanding what it is, let alone how to differentiate it from anything else. In fact the complete turnaround in acceptance, going completely overboard with labeling it everywhere instead of denying that it's a real thing, is probably one of the single worst mistakes the profession has ever done. They're still decades away from being able to do this safely.

It would actually be more productive to fund efforts to build a machine that could erase everything ever said or written about similar health problems, including from people's memories, to perform a complete reset from scratch than continuing on with stuff like this. Not that it's possible to build such a machine, but neither is achieving anything with the complete mess we have right now.

 

[Hutan]

Recent work utilizing objective measures indicates that depressed individuals spend more time engaged in sedentary
behaviors and less time daily engaged in light-intensity physical activity (e.g., walking) and moderate-to-
vigorous physical activity (e.g., exercise).15 Relatedly, sleep disturbances (either insomnia or hypersomnia) are common in depression

Actigraphy, specifically with the use of a triaxial accelerometer25 (e.g., Actigraph GT3X+), is a more reliable method of collecting activity data compared to self-report questionnaires Furthermore, actigraphy provides a more accurate detection of sleep behavior and duration compared to subjective measures of patient report in general populations and primary-mood disorder clinical populations.

Our sample comprised NHANES participants from collection cycles 2011-2012 and 2013-2014 (N = 8,378),41 including wrist-worn actigraphy information and depression scores (PHQ-9) (See Table 1). MDD presence was defined by a PHQ-9 composite score ≥ 10, a threshold which has been validated as an acceptable cut point for MDD detection in multiple studies42,43and meta analysis.44

Actigraphy data was collected via an Actigraph GT3X+, worn on participants’ non-dominant hands for atleast seven full days (midnight to midnight).39,40

The actigraphy data was then externally reviewed for data quality, and minute-level Monitor-Independent Movement Summary (MIMS) units were calculated.45 The present analyses used the minute-level triaxial MIMS value, which reflects the sum of the individual MIMS measurements obtained from the x-, y-, and z- axes, respectively.

 

[Hutan]

Theory-organized: Model Introspection

Using SHAP (Methods section 2.5.1), we found the following features to be the most influential to the model’s predictions for MDD presence: (1) Lower high-intensity activity across the full week (represented by 75th Quartile); (2) Higher high-intensity activity during the weekend (represented by 75th quartile); (3) Lower average activity across the full week; (4) A negative skew (left-skew) of weekend activity; (5) Lower high-intensity activity during the work week (represented by 75th quartile).

Self-organizing: General Population Model Performance

An LSTM model was implemented as a benchmark model, and performed only marginally above chance (AUCtest = 0.55, AUCvalidation = 0.52 ± 0.03). However, the modified “AlexNet” model showed the best performance across all approaches (AUCtest = 0.68, AUCvalidation = 0.63 ± 0.03) in detecting MDD in a general population.

On qualitative assessment, we note less defined boundaries between those times of typical sleep and wakefulness for individuals with MDD. In particular, we observe a more gradual movement increase in the morning (6AM-11AM) and a more gradual movement taper in the evening (6PM-12AM) compared to those without MDD.
In particular, we note areas of high influence in detecting MDD overnight (12AM-6AM), whereas the mid-morning (7AM-10AM) and evening (8PM-10PM) prove influential in detecting individuals without MDD.

That best model only had a sensitivity of 0.45 and specificity of 0.82.

They noted that simple regression models performed no better than chance.

 

[Hutan]

As such, we observed that depressed individuals are less active overall, with the largest difference in movement intensity occurring in the early-to-late morning.

In line with current literature,15,61 our findings provide support that less overall movement is related to higher depressive symptoms and provide important clinical implications. Existing treatments, including behavioral activation, target activity levels and depressive symptoms by implementing activity scheduling to improve one’s mood.66

Several digital interventions have been developed with a behavioral activation framework, making this treatment more accessible.67 Future research should investigate whether it would be beneficial for depressed individuals to utilize this intervention in the early morning, as we observed the largest differences in activity between the outcome groups during these hours.

Our work provides early support for the use of unobtrusive sensor data to detect MDD, which could then be acted upon by a tailored digital intervention.

It looks rather like people who have a job and have to get up in the morning are less likely to 'have depression'. And people who are awake at night are more likely to 'have depression'.

Of course, people with ME/CFS, and people with chronic pain are both more likely to be less active and to feel sad. It shouldn't follow from that that the solution to their depression is an app to 'make these people be more active'.