Endocrine Abstracts

Introduction: The prevalence of hepatic steatosis in people with obesity is very high. However, what is relevant is to detect those patients with associated fibrosis. The aim of the present study was to predict by artificial intelligence (AI) using clinical-analytical parameters, the presence of liver fibrosis in the biopsy performed during bariatric surgery (BS) of patients with obesity.

Subjects and methodology: Cross-sectional study of patients with obesity undergoing BS (2010-2018) who had liver biopsy (after informed consent) during the intervention. Patients with other liver diseases were excluded. Heat MAP was used initially to filter clinical and analytical variables with more association with the Target (fibrosis), subsequently the sample was randomly divided into 2 groups, 75% to train the AI models: Gradient-boosting, LGBM and xgboost. The remaining 25% to test efficacy and reduce overfitting. Sensitivity, specificity and accuracy were calculated using the best AI model.

Results: 362 patients (68.8% women), aged 47 [39-47] years, and BMI 42 [38-46] kg/m². Comorbidities: arterial hypertension 49.5%, Dyslipemia 46% and Type2 Diabetes (Dm²) 42%. Biopsies: 83.5% showed steatosis, 18% some degree of fibrosis (10.7%, 3.1% and 4.3% grades 1,2 and 3 respectively). It was observed in the heat map that the variables abdominal perimeter, Hypertension, Dm², Insulin, daily amount of alcohol ingested, Albumin, HBA1c, HOMA-IR index, C-peptide, HDLc, GOT, GPT and GGT, are more related to fibrosis. Gradient-boosting was the model that provided the best overall results. Among the variables finally used by the model, those with the greatest weight were GOT and HDLc, followed by GGT, HbA1c, Insulin, GPT, Albumin, amount of alcohol ingested, history of Dm² and HOMA-IR index. In the 25% of patients reserved for the accuracy test of the best model, it was observed that only 7.1% presented fibrosis in the biopsy, and the model predicted fibrosis in 12.5%, with only 1 false negative case. Finally, a Sensitivity:0.75, specificity:0.93 and accuracy of:0.827 was achieved.

Conclusions: The Gradient-Boosting model is highly used in medical studies since it can automatically identify more complex data structures, such as nonlinearity and diverse interactions; it has allowed to predict in previous studies cardiovascular events, delirium, sepsis, among o thers. In our sample it showed excellent specificity and good accuracy, which implies that its use in the clinic practice could avoid many liver biopsies, given the few false negatives it presents. This accuracy could even be improved with further training of the model