Large language model-assisted causal machine learning for identifying fatigue-related poor glycated hemoglobin in type 2 diabetes

BackgroundFatigue is common but mostly untreated in type 2 diabetes, since it requires a diagnostic workup which is hardly justified by fatigue alone. Individually identifying fatigue-related cause help decide effective follow-up and prioritize resource utilization. This study aimed to determine whether glycated hemoglobin (HbA1c) level contributes to fatigue for each individual case using a fatigue prediction model based on large language model (LLM)-assisted structural causal modeling (SCM). MethodsRetrospective cohort design was applied to collect data from diabetes management centers in Indonesia. We conducted SCM to select predictors among HbA1c and other variables for fatigue predictive modeling. Causal diagram was constructed by inferring the causal direction for each pair of correlated variables via LLM, i.e., GPT-4. The models were trained using eight machine learning (ML) algorithms. The best one was selected among the models that fulfilled sufficient sample size, was well-calibrated, and had positive net benefit using threshold closest to 95% specificity in our data. We chose the best model based on the area under curve (AUC) of receiver operating characteristics (ROC) and the concordance between the feature impact on model output based on the beeswarm plot of the Shapley additive explanation (SHAP) values and the effect size based on SCM. The SHAP waterfall plot was utilized to quantify HbA1c contribution to fatigue for each individual case. ResultsIndividuals with type 2 diabetes receiving OHA last 3 months (n=281) were more likely to report fatigue when they had poor HbA1c (adjusted odds ratio [aOR]=6.1, 95% CI 2.5, 14.4), comorbidity (aOR=25.2, 95% CI 12.1, 52.4), or a need for insulin treatment (aOR=3.6, 95% CI 2.0, 6.4). The best model used random forest algorithm (AUC-ROC=0.966, 95% CI 0.962, 0.969). Fatigue-related poor HbA1c could be individually identified among 95.1% (95% CI 94.5, 95.7) of those who reported fatigue. ConclusionsWe have developed a web application and nomogram for identifying fatigue-related poor HbA1c for each individual case. Future studies are warranted for external validation and randomized trials to examined the validity and impact of the cause identifier of fatigue in this study.