Predictive models for musculoskeletal injury risk: why statistical approach makes all the difference

Area of Science:

• Biostatistics
• Epidemiology
• Sports Medicine

Background:

• Injury prediction models are crucial for proactive health interventions in various populations.
• Previous models often dichotomized predictors, potentially oversimplifying complex relationships.
• The impact of predictor categorization on model performance requires further investigation.

Purpose of the Study:

• To compare the performance of an injury prediction model using categorized predictors versus continuous predictors.
• To evaluate the efficacy of selecting predictors based on univariate significance versus assessing non-linear relationships.
• To validate and replicate a previously developed injury prediction model in a new cohort.

Main Methods:

• A cohort of 1466 service members was followed for one year, collecting physical performance, medical history, and sociodemographic data.
• Four models were developed: original (dichotomized predictors), M2 (continuous predictors, linear assumption), M3 (continuous predictors, non-linear transformations), and M4 (clinically reasoned predictor selection).
• Model performance was assessed using R-squared, calibration (in the large and slope), discrimination, and decision curve analysis.

Main Results:

• Models M2 and M3, utilizing continuous and non-linear predictors, demonstrated substantially higher R-squared values (0.63-0.64) compared to the original model (0.07) and M4 (0.08).
• Calibration and discrimination metrics were superior in M2 and M3, indicating better accuracy and predictive power.
• Decision curve analysis showed significant net benefit improvements for M2 and M3 at risk thresholds of 0.25 and 0.50 compared to the original model.

Conclusions:

• Injury prediction models perform substantially worse when variables are dichotomized.
• Utilizing continuous predictors and assessing non-linear relationships significantly enhances model performance.
• Adherence to established recommendations for prediction model development is crucial for robust and accurate results.