EMA will be conducted via a smartphone app called Dear My Mind (currently under development; Figure 2). The EMA aims to capture real-time data on participants’ self-injury and suicidality, as well as mood appraisals. Participants will be prompted to complete the self-injury and suicidality questionnaire and the mood appraisal questionnaire 3 times a day for a duration of 14 days: at noon, in the evening, and at night (Table 1). Each morning, participants will be asked to answer a questionnaire regarding their sleep patterns from the previous night. This questionnaire will provide additional information about participants’ sleep conditions to complement the objective sleep measures collected. Furthermore, participants will be asked to complete a mood diary each night (the Mood Diary section below presents detailed information).

The visual analog scale (VAS) is a rating of the participant’s perceived level of “self-injurious & suicidal thoughts and urges.” It is constructed on a horizontal line with “not at all” at one end and “very much” at the other (on a 7-point Likert scale). It can quantify a participant’s self-injurious and suicidal thoughts and urges and is useful for capturing changes within individuals rather than differences between individuals. In addition, if the participant has attempted self-injury since the last response, a text box prompts the participant to indicate the method of the attempt. There are also additional questions about antecedent events, antecedent emotions that occurred within the 24 hours prior to the self-injury, and experiences immediately following the self-injury. If the participant indicates that they have attempted suicide since their last response, they will be classified as high risk and the monitoring researcher will contact them immediately to refer them to a local mental health clinic or counseling center for immediate help.

Participants will be instructed to respond to the VAS 3 times a day about the emotions they are experiencing at that moment. The mood VAS will be constructed on a horizontal line with “not at all” at one end and “very much” at the other (on a 9-point Likert scale). Specifically, participants will be asked about their feelings of depression, sadness, anxiety, fear, loneliness, rejection/hurt, anger at self, anger at others, shame, and emptiness.

Participants will be asked to engage in daily mood diary entries using a free-text box. Each night, participants will be prompted to write about an event they experienced during the day, how it affected their mood, and the strategies they used to resolve it. Participants receive clear instructions regarding the purpose and content of the mood diary entries. A free-text format provides participants the flexibility to express their thoughts and feelings in their own words, allowing for a more nuanced understanding of their mood dynamics. Additionally, we emphasize the importance of consistent and honest reporting to maximize the validity and reliability of the data collected. The purpose of the mood diary is to capture participants’ subjective experiences and emotions, as well as their efforts to manage and overcome daily challenges. Participants will be encouraged to be as detailed and specific as possible when describing the event, their emotional response, and their coping strategies. This information will provide valuable insights into the contextual factors that influence participants’ mood fluctuations and their adaptive responses. The daily nature of the mood diary entries allows for the examination of day-to-day variations in mood and the exploration of the relationship between specific events and emotional well-being. It provides a comprehensive picture of participants’ subjective experiences and contributes to the overall understanding of their digital phenotypes.

To complement the objective sleep data obtained from the wearable device, participants will also be administered the sleep condition questionnaire to gather subjective sleep measures. This questionnaire includes items related to bedtime and wake time, actual sleep duration, and subjective sleep quality. Participants will be asked to respond to the sleep condition questionnaire once a day, specifically each morning, providing information about their sleep experience during the previous night. By capturing participants’ subjective perceptions of their sleep, this questionnaire will help elucidate participants’ personal experiences and perceptions of their sleep quality and add a valuable subjective component to the objective sleep data collected. Participants will be encouraged to provide accurate and detailed responses to provide a comprehensive understanding of their sleep patterns and experiences. By combining objective sleep data from the wearable device with participants’ subjective reports, this study will gain a more comprehensive and holistic understanding of participants’ sleep conditions.

Initially, basic data analysis techniques such as 2-tailed independent t tests, linear regression analysis, repeated measures ANOVA, or other appropriate methods will be used to analyze the data and identify any significant changes or associations.

To examine dynamic changes at the within-individual level, dynamic structural equation modeling (DSEM) will be used. This method allows for the investigation of time-varying relationships and the identification of individual-level patterns of change over time. Specifically, we will examine whether an individual’s affective state, physiological state (heart rate, heart rate variability), and sleep state predict NSSI thoughts and behaviors from one observation window to the next. To address these 2-level structured data and to examine the temporal relationships between predictor variables and NSSI thoughts and behavior, multilevel vector autoregressive models were constructed within the framework of DSEM. This approach will enable us to investigate the extent to which time-varying variables at t–1 (eg, depressive affect) predict NSSI thoughts and NSSI behavior at t.

To comprehensively analyze participants’ digital phenotypes and predict NSSI thoughts and behaviors, sparse logistic regression will be used. Logistic regression predicts the probability of being categorized into 1 of 2 groups when the response variable is binary. This model has the useful property that the estimated coefficients are log odds ratios, but the results are difficult to interpret when the number of variables becomes large. To overcome this problem, sparse logistic regression using the least absolute shrinkage and selection operator (LASSO) can be considered. This model simultaneously conducts feature selection and estimation, allowing for interpretation with a few selected important predictors. To evaluate prediction performance, we will divide the data set into a 30% subset for model estimation and a 70% subset for model training. Additionally, we will perform 5-fold cross validation to avoid overfitting the model. In this process, the full data set will be divided into 5 subsets, with 4 subsets used for training and the remaining subset used as an independent validation set. To solve the problem of imbalance in the training data, the synthetic minority oversampling technique (SMOTE) will be used. SMOTE is an oversampling technique that takes samples from a small number of classes in the data and adds random values to generate new samples to add to the data. This algorithm helps to address the issue of the overfitting problem caused by random oversampling [52].

Measures including accuracy, precision, sensitivity, specificity, and area under the curve will be calculated to explore the prediction performance of sparse logistic regression.

The outcome variables to be entered into the model are NSSI thoughts and behaviors, which will be coded dichotomously (0: absent; 1: present). The predictor variables to be included in the model as interval averages are step count, heart rate, and heart rate variability. These variables will be binned according to the intervals at which the dependent variable, NSSI thoughts and behaviors, will be assessed. The mean value per bin will then be calculated and used as a predictor. For example, as active data collection occurs 4 times, at 9 AM, noon, 6 PM, and 9 PM, the numerous values in the passive data will be divided into 4 bins: 8:00 AM to 11:59 AM, noon to 5:59 PM, 6:00 PM to 8:59 PM, and 9:00 PM to 23:59 PM. For heart rate, values less than 30 or greater than 200 beats per minute (bpm) will be discarded. Heart rate will be calculated as the mean heart rate in bpm, and heart rate variability will be obtained via the SD of the normal-to-normal interval (SDNN), which is the SD of the NN interval and the square root of the variance. Subtracting the mean from each heart rate interval yields the SD of heart rate intervals. Regarding step count, the total step count, average, median, and SD for each period will be calculated. For sleep data, preprocessing will ensure only 1 value per day, representing 1 value per night. Sleep data preprocessing will include the average, median, and SD of core sleep, deep sleep, REM sleep, and total sleeping duration. The ratio of each sleep duration to the total sleep duration will also be calculated as a reference for sleep quality.