Please rotate your device to landscape mode to view the charts.

Mortality prediction using data from wearable activity trackers and individual characteristics: An explainable artificial intelligence approach

Journal: Expert Systems with Applications

Date: 2025

Author: Byron Graham, Mark Farrell

Abstract:
Mortality prediction plays a crucial role in healthcare by supporting informed decision-making for both public and personal health management. This study uses novel data sources such as wearable activity tracking devices, combined with explainable artificial intelligence methods, to enhance the accuracy and interpretability of mortality predictions. By using data from the UK Biobank—specifically wrist-worn accelerometer data, hospital records, and various demographic and lifestyle factors, and health-related factors—this research uncovers new insights into the predictors of mortality. Explainable artificial intelligence techniques are employed to make the models’ predictions more transparent and understandable, thereby improving their practical applications in healthcare decisions. Our analysis shows that random forest models achieve the highest prediction accuracy, with an area under the curve score of 0.78. Key predictors of mortality include age, physical activity levels captured by accelerometers, and other health and lifestyle factors. The study also identifies non-linear relationships between these predictors and mortality, and provides detailed explanations for individual-level predictions, offering deeper insights into risk factors.

Link: Google Scholar

Background and Context

Research Purpose

This study examines how data from wearable activity trackers combined with explainable artificial intelligence can enhance mortality predictions to support healthcare decision-making.

Data Sources

The research analyzed data from 96,695 UK Biobank participants including wrist-worn accelerometer readings, hospital records, demographic factors, and lifestyle characteristics.

Methodology

Seven machine learning algorithms were compared for predicting 3-year mortality, with random forest models achieving highest accuracy and explainable AI techniques used to interpret results.

Random Forest Model Achieves Highest Prediction Accuracy Among ML Algorithms

  • Random Forest algorithm achieved the highest accuracy with an AUC score of 0.783
  • All tested machine learning models performed better than traditional statistical approaches
  • Model accuracy is crucial for reliable mortality predictions in healthcare applications

Most Important Predictors of Mortality Based on SHAP Values

  • Age was found to be the most important predictor of mortality
  • Physical activity (measured by accelerometer) was the second most important factor
  • Health conditions and socioeconomic factors also played significant roles

Relationship Between Physical Activity and Mortality Risk

  • Higher levels of physical activity were associated with decreased mortality risk
  • The relationship shows diminishing returns at very high activity levels
  • Objective measurement through accelerometers provides more reliable activity data than self-reporting

Impact of Recent Hospital Episodes on Mortality Risk

  • Recent hospital episodes significantly increased mortality risk
  • 16.2% of patients with recent hospital episodes died within 3 years
  • Hospital episode data provides valuable predictive information for mortality risk assessment

Age-Related Changes in Mortality Risk

  • Mortality risk increases significantly after age 60
  • The relationship between age and mortality is non-linear
  • Age remains the strongest predictor even when accounting for other factors

Contribution and Implications

  • The study demonstrates how wearable device data can improve mortality predictions when combined with traditional health metrics
  • Findings can be applied to develop more accurate risk assessment tools for healthcare and insurance applications
  • The explainable AI approach provides transparency in how predictions are made, building trust in the model's decisions

Data Sources

  • Model accuracy comparison (Finding 1) based on Table 3 in the article
  • Predictor importance (Finding 2) derived from SHAP values shown in Figure 2
  • Physical activity relationship (Finding 3) based on partial dependence plots in Figure 4
  • Hospital episode impact (Finding 4) calculated from data in Table 2
  • Age-related mortality risk (Finding 5) based on patterns described in results section and Figure 4