[[sensors-20-03572.pdf]] - Nicholas C Jacobson # ‪Passive Sensing of Prediction of Moment-to-Moment Depressed Mood among Undergraduates with Clinical Levels of Depression Sample using Smartphones‬ 2020 paper. **Question: Can we detect or predict patients' moods using passive monitoring without their active input?** ==My Takeaway== 1. Passive Sensors 2. Mood EMA - every hour 3. Algorithm takes the last 24 hours of data to predict the next 1-hour mood. 4. When comparing Observed (patient input) with Predicted Mood - There was a significant, positive, strong correlation ( r=0.587) Limitation: Participants are undergrads, not MDD diagnosed patients. Support from existing study: *Existing research tends to find a significant relationship between predicted depressive symptoms using passive sensor data and actual symptom severity: 1. Mobile phone location data across ten weeks 2. Mobile sensing and support 8 weeks 3. smartphone data 12 weeks. 4. 8 weeks smartphone based monitoring 5. bipolar patients 12 weeks, inertial sensors and GPS 6. 13 weeks sensor and wearable 7. 9 days smartphone senors 8. 3 and 6 weeks mobile phone sensors* **Is it essential to track mood fluctuations or persistent low mood? The following paragraph said that MDD symptoms occur rapidly** - Importantly, shifts in MDD symptoms occur rapidly with substantial fluctuations occurring over the course of a day or even hour-to-hour [32–35]. Consequently, it is essential to predict MDD symptoms across intervals as short as hours. **The use of Geolocation data can be correlated with depressive severity?** [[DP - Geolocation]] #[[202202151536 The use of Geo-location to trigger EMI, to remind user about their training, to reduce relapses]] - Canzian and Musolesi (2015) predicted daily mood by examining daily location data from smartphone sensors, finding that geolocation data was correlated with depression severity on a day-to-day basis and that geolocation data could be used to predict dichotomous depression severity from 1 to 14 days later. There are limitations to prior studies. 1. **Prior studies only look at whether patient is depressed or not depressed instead of the level of depression.** 1. Prior studies dichotomized their depression outcomes when examining their primary outcomes, rather than looking at whether depressed mood could be predicted across a continuum 2. **Did not consider how mood fluctuates even within a single day. Mood changes can occur rapidly.** *This reminds me of the [[HOPES - Components for the EMA EMI Design]], [[Designing EMI and EMA]] - Importance of multiple EMA at random timing.* 3. **If you want to help someone on a rainy day, give the umbrella at the right time.** 1. Secondly, two of the three studies did not examine depressed mood within days, which neglects the substantial mood changes in MDD across a single day [32], rapid mood fluctuations on an hourly basis [33,35,36], and a great deal of variation in depressed mood not being stable across more than several hours [37]. 2. Consequently, most prior research is unable to adequately translate to inform just-in-time adaptive interventions (JITAI) [38,39], as this research might miss important times in which persons might be experiencing depressed mood fluctuations. ## Measures - Baseline Depression Severity - Depression Anxiety and Stress- Depression Scale - 14 Item self-administrated questionnaire. - Dynamic Depressed Mood - Positive and Negative Schedule Expanded (PANAS-X) - Participants were asked once per hour. To rate a 100-point scale, the extent they feel "sad" and "lonely" at that moment. - "Loneliness is strongly linked to MDD [54-57]" - This passage shows support showing how "Sad" and "Lonely" are linked to depression. - “In particular, it may be worthy to note that the two items were more strongly associated with depression than the “depressed” item itself (which had a coefficient of 0.49), suggesting that self-reported sadness and loneliness may be strong predictors of depression (thus corroborating/justifying the current study’s use of the “sad” and “lonely” items to assess dynamic depressed mood). Furthermore, another study demonstrated strong convergent validity (r = 0.66 to 0.67) between sadness and loneliness [59].” (“‪Passive Sensing of Prediction of Moment-to-Moment Depressed Mood among Undergraduates with Clinical Levels of Depression Sample using Smartphones‬”, p. 4) - [[Calculation of Coefficient]] - Higher scores on the scale indicate a greater depressed mood. ## Passive Sensor Data - 1) direct location based information: (1a) GPS coordinates (latitude, longitude), (1b) location accuracy, (1c) location speed, and (1d) whether the location-based information was based on GPS or WiFi; (2) location type based on the Google Places location type (e.g., University, gym, bar, church); (3) local weather information, including (3a) temperature, (3b) humidity, (3c) precipitation, (2) light level, (3) heart rate information: (3a) average heart rate and (3b) heart rate variability; and (4) outgoing phone calls. ## Planned Analysis - **They use the last 24 hours' data to predict the level of depression in a patient in the next 1 hour** - The goal of the modelling strategy is to use the past 24 hour sensor data to predict the next hour change of depression symptom severity based on the passive data from the next hour. - What is the general pattern ?? "nomothetic models focus on general patterns across groups of people, while idiographic approaches focus on the unique details of individuals." - **The researchers is able to predict mood changes and also intraindividual variability; different people's mood changes differently.** - In addition to being interested in predicting depressed mood across all conditions, we were also interested in predicting only intraindividual variability within each person (see Figure 4). The results suggested that there was significant intraindividual variability predicted with an average correlation of 0.376, 95% CI [0.226, 0.508].