Key points are not available for this paper at this time.
Objective The current research aims to utilize a fine-tuning-free, freely available large language model (LLM) framework to systematically investigate the correlation between the platelet-to-albumin ratio (PAR) and the pathological severity of diabetic nephropathy (DN). This approach aims to develop an exploratory risk stratification tool for this microvascular complication. Methods A retrospective analysis was conducted on 195 patients diagnosed with DN. The effectiveness of the PAR as a diagnostic tool was evaluated. A “tabular-to-text-to-vector” framework was developed, utilizing the frozen Qwen and Llama models to extract semantic features from the serialized clinical narratives. The performance of this model was compared with traditional algorithms using a 5-fold cross-validation technique. Results Receiver operating characteristic (ROC) analysis of the current data indicated that the optimal cutoff for the PAR was 7.155, with an area under the curve (AUC) of 0.716. Multivariate logistic regression analysis revealed a positive correlation between PAR levels and the pathological severity of DN (OR: 6.65, 95% CI: 2.617–16.9). The fusion LLM framework showed improved balance in addressing class imbalance achieving higher specificity (56.67%) compared to the random forest (RF) model (31.67%), although improvements in overall AUC were marginal. Regarding the assessment of interstitial fibrosis and tubular atrophy (IFTA), the model's macro-F1 score of 51.00 ± 5.71%, exceeded that of the XGBoost model, which recorded a score of 45.22%. Conclusion The platelet-to-albumin ratio was significantly associated with the pathological severity of DN. As an exploratory proof-of-concept, the fine-tuning-free fusion LLM framework proposed in this study utilizes semantic reasoning and demonstrates potential scalability for small-sample medical datasets, particularly in non-invasive risk stratification.
Xia et al. (Wed,) studied this question.