Interpretable machine learning models for detecting peripheral neuropathy and lower extremity arterial disease in diabetics: an analysis of critical shared and unique risk factors

Ya Wu, Danmeng Dong, Lijie Zhu, Zihong Luo, Yang Liu, Xiaoyun XieBMC Medical Informatics Decis. Mak 2024 · DOI

TL;DR

This study developed interpretable machine learning models to predict diabetic peripheral neuropathy (DPN) and lower extremity arterial disease (LEAD) using readily available clinical and lab data. The models identified key shared risk factors like elevated UACR and HbA1c, as well as unique predictors for each condition, offering a promising avenue for early detection and prevention of diabetic foot ulcers.

What Problem Does This Paper Address?

Diabetic peripheral neuropathy (DPN) and lower extremity arterial disease (LEAD) are serious complications of diabetes that significantly increase the risk of diabetic foot ulcers (DFUs). DFUs can lead to amputations, increased mortality, and substantial economic burdens. Early identification and management of DPN and LEAD are crucial for preventing these severe outcomes. However, current diagnostic methods often rely on specialized equipment and trained personnel, which are not always accessible, especially in resource-limited regions. While machine learning (ML) has shown promise in predicting these conditions, most studies focus on either DPN or LEAD individually, neglecting the potential for shared risk factors and a unified approach. This paper aims to bridge this gap by developing ML models for both DPN and LEAD simultaneously and identifying both common and distinct risk factors.

Methods

The researchers employed a retrospective study design, collecting clinical and laboratory data from 479 diabetic inpatients. To build predictive models, they utilized three machine learning algorithms: Logistic Regression (LR), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost). To ensure the models were robust and avoided overfitting, they implemented a multi-pronged feature selection strategy using:

  • Mutual Information (MI): Measures the statistical dependency between variables.
  • Random Forest Recursive Feature Elimination (RF-RFE): Iteratively removes less important features.
  • Boruta Algorithm: Compares feature importance against randomized "shadow" features.

Features selected by at least two of these methods were used for model development. To further optimize model performance, Particle Swarm Optimization (PSO) was used for hyperparameter tuning. Crucially, to understand why the models made their predictions, the SHapley Additive exPlanations (SHAP) method was applied. This technique provides a game-theoretic approach to explain the contribution of each feature to the model's output, allowing for the identification of critical risk factors. For the LEAD dataset, which had an imbalanced class distribution, a random undersampling technique was used to create a more balanced training set.

Datasets

The study utilized a dataset of 479 diabetic inpatients from Tongji Hospital, collected between January 2022 and March 2023. This dataset included demographic information, clinical characteristics, and laboratory results. The data was split into an 80% training set and a 20% testing set for model development and evaluation. No new benchmark datasets were introduced; the study used existing patient data.

Key Insights and Ideas

  • Unified Approach for DPN and LEAD: The study successfully developed ML models for both DPN and LEAD, highlighting the potential for a more integrated approach to managing these interconnected diabetic complications.
  • Interpretability is Key: The use of SHAP values is a significant contribution, moving beyond "black box" ML models to provide actionable insights into the specific factors driving predictions. This is crucial for clinical adoption.
  • Shared Risk Factors are Powerful Targets: Identifying common risk factors like elevated UACR, HbA1c, Scr, advanced age, and carotid stenosis suggests that interventions targeting these areas could simultaneously reduce the risk of both DPN and LEAD, thereby lowering the overall risk of DFUs.
  • Unique Predictors Offer Nuance: The discovery of distinct risk factors (e.g., low serum albumin for DPN, high NLR for LEAD) allows for more tailored management strategies for individual patients.
  • Accessible Data for Prediction: The reliance on standard clinical and laboratory data makes these models highly practical for implementation in various healthcare settings, including those with limited resources.

Key Results

  • DPN Prediction: The XGBoost model demonstrated superior performance for DPN detection, achieving an AUC of 0.903, recall of 83.7%, specificity of 86.8%, and accuracy of 85.4%.
  • LEAD Prediction: The Random Forest model excelled in predicting LEAD, with an AUC of 0.923, recall of 85.7%, specificity of 92.9%, and accuracy of 91.9%.
  • Shared Risk Factors: The top five critical risk factors common to both DPN and LEAD identified by SHAP analysis were: increased urinary albumin-to-creatinine ratio (UACR), glycosylated hemoglobin (HbA1c), serum creatinine (Scr), advanced age, and carotid stenosis.
  • Unique Risk Factors:
    • DPN: Decreased serum albumin and lower lymphocyte count were uniquely associated with DPN.
  • LEAD: Increased neutrophil-to-lymphocyte ratio (NLR) and higher D-dimer levels were uniquely associated with LEAD.

Limitations and Future Work

The authors acknowledge that this was a cross-sectional study, which limits the ability to establish causality. Future work should focus on:

  • Prospective Studies: To validate these findings and establish causal relationships.
  • Broader and More Diverse Populations: Expanding the study to include a wider range of patients to enhance generalizability.
  • Unified ML Model: Investigating the feasibility of developing a single ML model that can accurately predict both DPN and LEAD simultaneously.
  • Personalized Treatment Strategies: Leveraging the identified risk factors to develop more personalized and comprehensive treatment plans.

Summary updated

BibTeX
@article{wu2024interpretable,
  author = {Ya Wu and Danmeng Dong and Lijie Zhu and Zihong Luo and Yang Liu and Xiaoyun Xie},
  title = {{Interpretable machine learning models for detecting peripheral neuropathy and lower extremity arterial disease in diabetics: an analysis of critical shared and unique risk factors}},
  journal = {BMC Medical Informatics Decis. Mak.},
  year = {2024},
  doi = {10.1186/s12911-024-02595-z},
}
Figure