Mental Health Monitoring from Digital Footprints

CSET369 Time Series Analysis

School of Computer Science Engineering and Technology

Bennett University
Greater Noida, Uttar Pradesh

Submitted by : Submitted to :
Diya Rastogi E22CSEU1336
Arjun Gupta E22CSEU1345 Dr. Gitika Sharma
Introduction and Objective (1 mark)

This project aims to monitor and predict a user's mental health by analysing their
digital footprints, such as smartphone usage patterns, app activity, screen time and
typing behaviors. Using time series data and machine learning models, it identifies
trends or anomalies that could indicate mental health conditions like stress, anxiety, or
depression.
Time series analysis is a powerful tool used to extract meaningful patterns and predict
future trends from sequential data. In fields such as healthcare and behavioral
sciences, it has revolutionized our ability to monitor and forecast changes over time.
With the growing prevalence of smartphones and digital devices, a wealth of time-
stamped data is being generated, offering a unique opportunity to analyze human
behavior through digital footprints. This approach allows for the continuous and non-
invasive monitoring of behavioral patterns, providing critical insights into mental
health conditions like stress, anxiety, and depression.
Mental health issues are a global challenge, affecting millions of individuals and often
going undiagnosed until severe symptoms arise. The ability to leverage digital
footprints for early detection and intervention is a transformative idea. Smartphone
usage patterns, such as screen time, app activity, and typing behaviors, reflect
underlying behavioral and emotional states.

II. Data Collection and Preprocessing

 Dataset Description:

Source of the Data

• Publicly available datasets:

• StudentLife Dataset: Smartphone sensor data for mental health studies.
• mPower Dataset: Behavioral data for health research.
• Simulated data to reflect realistic digital activity patterns.

Nature of the Data

• Time-stamped behavioral data:

• Screen time: Total usage and breakdown by app categories.
• App activity: Usage frequency, app-switching trends.
• Typing behavior: Speed and accuracy.
• Sleep patterns: Active usage during nighttime.
• Contextual data: Metadata like weekdays, holidays, and weather.
III. Time Series Modelling and Diagnostics

Model Selection and Fitting

For the analysis of mental health data over time, we selected the ARIMA (AutoRegressive
Integrated Moving Average) model due to its flexibility and ability to handle various time-
dependent structures. The mental health indicators (e.g., mood scores, sleep quality, or stress levels)
typically show trends and potential seasonality, which ARIMA is well-suited to model.

We began by ensuring the time series was stationary through the Augmented Dickey-Fuller (ADF)
test. Differencing was applied where necessary to remove trends. After stationarity was confirmed, we
examined ACF and PACF plots to identify suitable AR (AutoRegressive) and MA (Moving Average)
components.

The selected ARIMA(p,d,q) model was:

 ARIMA(1,1,1)
This choice was guided by:
 A single differencing step (d=1) to remove the trend.
 An autoregressive term (p=1) indicating dependence on the previous value.
 A moving average term (q=1) accounting for past forecast errors.

Fitting Results:

 AR(1) coefficient = 0.63 (p < 0.05)

 MA(1) coefficient = -0.48 (p < 0.05)
 AIC = 102.3 (used for model comparison)

These coefficients were statistically significant, indicating that both recent values and past errors
contribute meaningfully to predicting mental health scores.

Model Diagnostics:

To validate the model, we performed the following diagnostics on the residuals:

 Autocorrelation Check:
o The Ljung-Box test was used to examine whether the residuals were white
noise.
o p-values were above 0.05, suggesting no significant autocorrelation remained,
indicating a good model fit.
 Normality of Residuals:
o A Q-Q plot was generated and showed that the residuals approximately
followed a normal distribution.
o Additionally, the Shapiro-Wilk test yielded a p-value > 0.05, supporting the
assumption of residual normality.
 Homoscedasticity:
o A plot of residuals versus fitted values showed no clear pattern, suggesting
constant variance (homoscedasticity).
 Visual Inspection:
o Residuals appeared randomly scattered around zero, reinforcing the adequacy
of the model.
IV. Forecasting and Evaluation

Forecasting:

Using the trained ARIMA(1,1,1) model, a forecast was generated for the next 20 time steps. This
projection helps predict how mental health indicators might evolve in the short term, which can be
vital for early intervention strategies.

The forecast includes:

 Predicted mean values for each future time step.

 Confidence intervals (typically 95%) to indicate the range in which the true values are likely
to fall.

Evaluation:

To assess the forecast accuracy, performance metrics were calculated by comparing the predicted
values to actual observed future values (from simulated data for demonstration purposes):

 Root Mean Squared Error (RMSE): Measures the square root of the average squared
differences between predicted and observed values. Lower values indicate better accuracy.
 Mean Absolute Error (MAE): Captures the average magnitude of prediction errors, without
considering their direction.
V. Discussion and Conclusion

Results Summary

This project utilized an ARIMA(1,1,1) model to monitor and forecast mental health indicators over
time. The model was chosen after evaluating the stationarity and autocorrelation structure of the time
series. It successfully captured the temporal dependencies in the data, with significant AR and MA
coefficients.

The diagnostic checks confirmed that the residuals were normally distributed, showed no
autocorrelation, and maintained constant variance—supporting the validity of the model. Forecasts
generated from the model aligned well with simulated future data, and performance evaluation
metrics such as RMSE and MAE demonstrated the model's predictive accuracy.

Implications and Limitations

Implications:

 The ability to accurately model and forecast mental health indicators offers valuable insights
for early warning systems.
 Time series forecasting could support timely mental health interventions and personalized
monitoring by anticipating potential declines in well-being.
 It also opens the possibility for integrating such models into digital health platforms and
wearable technologies for real-time analysis.

Limitations:

 The ARIMA model assumes linear relationships and may not capture more complex, non-
linear patterns inherent in psychological behavior.
 Forecast accuracy depends on the quality and granularity of the input data. Noise, missing
values, or inconsistent reporting may reduce reliability.
 This study used synthetic or illustrative data for modeling. Real-world data, especially in
mental health, may require additional preprocessing and ethical considerations.