Prediction of type 2 diabetes mellitus onset using logistic regression-based scorecards

Yochai Edlitz, Eran SegaleLife 2022 · DOI

TL;DR

This paper introduces two accessible, scorecard-based models for predicting type 2 diabetes (T2D) onset: one using simple anthropometric measures and another using four common blood tests. These models outperform existing risk scores and offer a practical tool for early risk assessment, potentially enabling timely interventions and lifestyle changes.

What Problem Does This Paper Address?

Type 2 diabetes (T2D) is a global epidemic with significant health and economic consequences. Early detection of individuals at high risk is crucial for prevention through lifestyle modifications or medication. However, existing prediction models can be complex, require specialized equipment, or are not universally accessible, particularly in lower socio-demographic populations. This paper aims to bridge this gap by developing accurate, yet simple and accessible, T2D risk prediction models.

Methods

The researchers employed a multi-stage approach to develop and validate their T2D prediction models. They began by analyzing a large dataset from the UK Biobank, starting with 798 potential features. Through feature selection processes, including gradient boosting decision trees and survival analysis, they identified key predictors.

The core innovation lies in developing two simplified, scorecard-based models:

  1. Non-laboratory Anthropometric Model: This model uses easily obtainable measurements like sex, age, weight, height, waist circumference, hip circumference, waist-to-hip ratio, and body mass index.
  2. Laboratory Blood Test Model: This model is more precise and incorporates age, sex, and four common blood tests: high-density lipoprotein (HDL), gamma-glutamyl transferase (GGT), glycated hemoglobin (HbA1c%), and triglycerides.

Both models were simplified into an accessible scorecard format. They also employed techniques like iterative feature elimination, cross-validation, and bootstrapping for model optimization and calibration. External validation was performed using the electronic medical record database of Clalit Health Services in Israel.

Datasets

  • UK Biobank (UKB): This observational study cohort provided the primary dataset for training and validation. The study included 502,536 participants aged 40-69, with a subset of 44,709 participants used for model development.
  • Clalit Health Services: This Israeli electronic medical record database served as an external validation cohort to assess the generalizability of the four blood test model.

Key Insights and Ideas

  • Simplicity Trumps Complexity: The study demonstrates that highly accurate T2D prediction can be achieved with relatively simple, accessible features, challenging the notion that complex models with numerous features are always superior.
  • Four Blood Tests are Powerful: A combination of just four common blood tests, along with age and sex, proved to be a highly effective predictor of T2D onset, outperforming many other feature combinations.
  • Anthropometrics as a Practical First Step: The anthropometric scorecard offers a valuable, non-invasive tool for initial risk stratification, especially in settings where laboratory tests are not readily available.
  • HbA1c% Dominance: As expected, HbA1c% emerged as the most predictive feature in the blood test model, highlighting its critical role in T2D diagnosis and risk assessment.
  • Age and Sex Less Critical with Blood Tests: Interestingly, age and sex showed low feature importance in the blood test model, suggesting their predictive information might be implicitly captured by the blood test markers themselves.

Key Results

  • The non-laboratory anthropometric scorecard achieved an auROC of 0.81 (95% CI 0.77–0.84), classifying participants into three risk groups with predicted T2D risks of 1%, 5%, and 9%.
  • The four blood test scorecard model achieved a higher auROC of 0.87 (95% CI 0.85–0.90), stratifying participants into four risk groups with predicted T2D risks of <0.5%, 3%, 10%, and a high-risk group of 23%.
  • The four blood test model's performance on the external Clalit cohort was an auROC of 0.75 (95% CI 0.74–0.75), demonstrating its general applicability, though slightly lower than in the UKB cohort.
  • Both developed models outperformed established risk scores like FINDRISC (auROC 0.73) and GDRS (auROC 0.66) in the UKB cohort.

Limitations and Future Work

  • Cohort Bias: The study acknowledges that the T2D prevalence in their UKB cohort (1.79%) is lower than the general UK population, likely due to "healthy volunteer selection bias" and "attrition bias." This might affect the generalizability of the models to broader populations without further fine-tuning.
  • External Validation Nuances: While the Clalit cohort provided external validation, its retrospective nature and different patient selection criteria (higher T2D prevalence, more targeted blood testing) led to slightly lower performance compared to the UKB cohort.
  • Future Work: The authors suggest that applying their models to additional populations and fine-tuning feature coefficients may be necessary for broader applicability. They also highlight the potential for these models to serve as personal catalysts for lifestyle changes.

Summary updated

Abstract

The funders had no role in study design, data collection, interpretation, or the decision to submit the work for publication.
BibTeX
@inproceedings{edlitz2022prediction,
  author = {Yochai Edlitz and Eran Segal},
  title = {{Prediction of type 2 diabetes mellitus onset using logistic regression-based scorecards}},
  booktitle = {eLife},
  year = {2022},
  doi = {10.7554/elife.71862},
}
Figure