Machine learning for diabetes prediction and personalized treatment recommendations after acute pancreatitis

Study cohort and baseline characteristics

Between 1 October 2016 and 31 October 2021, 3477 admissions in AP were screened with the Hospital Information System (HIS).After using the exclusion criteria described (Supplementary Figure 1), 820 AP patients without known diabetes were included in our study. Of these, two-thirds (n = 574) were randomly assigned to the training set and the remaining one-third of his (n = 246) were assigned to the validation cohort.table 1 Indicates patient baseline characteristics. The median age was her 50 years (38, 63 years). The percentage of males he was 61.3% (n = 503). The biliary tract was the most common cause of AP. 484 (59%) patients had mild AP, 280 (34.1%) moderate AP and 56 (6.8%) severe AP. Sixty-eight (8.3%) patients had PPDM-A, were likely obese (20.7% vs. 9.2%, P = 0.005), presented with hyperlipidemia, and had nonalcoholic fatty liver disease (NAFLD). ) tended to merge (75% vs 45.2%, P < 0.001). Smoking rates were higher in her PPDM-A patients than in those without DM.

feature extraction

Lasso regression (L1 regularized logistic regression) can be used for feature extraction in classification models. We performed 1000 random perturbed lasso regressions to extract the weights of the 38 clinical features. We ranked the average weights of these 38 features and used a threshold of 0.01 to obtain the 9 measures that most affected the model’s classification (Fig. 1), admission blood sugar, and obesity (in order of BMI > 28 kg/m)²), cardiovascular disease (CVD), age, NAFLD, alanine transaminase (ALT), uric acid (UA), HDL-C < 1.03 mmol/l, smoking. In addition, the remaining indices with a range greater than 0 included several features such as alcohol consumption, organ failure, acute peripancreatic fluid retention (APFC), blood urea nitrogen (BUN), creatinine, hypertension, amylase, and Ca. It was The results indicate that the two most influential factors in PPDM-A remain admission blood glucose levels and obesity. These two indicators are also indicators related to type 2 diabetes. This suggests that type 2 diabetes and his PPDM-A share common risk factors.

Screening for core influencing factors. Mean feature weights were ranked by 1000 lasso regressions. Among these, 9 features with Feature Importance Score >0.01 were selected as core genes thought to be related to PPDM-A.

Algorithm performance

Multiple machine learning algorithms were used to build the classification model. Following the same approach, we built a classification model based on the core 9 features. We validated the model’s performance on the training set using 5-fold cross-validation (Figure 1). 2A, B; Supplementary Table 1). Additionally, we performed internal validation on the training set (Figure 2). 2CD; Supplementary table 2). We then tested these models with validation data (Fig. 2E, F; Supplementary Table 3, Four). The results showed that LR L1(C = 1) gave the best model at the average level (AUC = 0.819, CA = 0.927, F1 = 0.912, Precision = 0.912, Recall = 0.927). 2E, Supplementary table 3). For the prediction of positive events, LR L1(C = 1) also achieved the best results (AUC = 0.819, CA = 0.927, F1 = 0.357, Precision = 0.625, Recall = 0.250; Fig. 2F, Supplementary table Four). Previous analyzes showed that prognostic models built using the core 9 features had the best predictive effect.

model performance. Five-fold cross-validation was used to evaluate model performance on the training set. ROC curves and calibration curves were used to compare the strengths and weaknesses of the models. (a,B.) ROC and oscillation curves of the five machine learning models in the training set using 5-fold cross-validation. (C.,D.) internal validation of the training set. (picture,debt) ROC and oscillation curves of the five machine learning models in the test set.From the supplementary table Fourwe find that the model obtained by logistic regression (L1 regularization) performs best with AUC = 0.819 and F1 = 0.357 on the validation set.

Assessing the interpretability of model predictions

A nomogram was constructed based on LR L1(C = 1) for the 9 features. HDL-C < 1.03, CVD, and ALT were predicted to contribute to PPDM-A (Fig. 3A) Negative. To gain insight into the features that contributed most to the model’s predictive results, we used Sharply Values ​​to assess the importance of core features for model evaluation (Fig. 3B). Factors that most influenced model predictions included HDL-C < 1.03 mmol/l、BMI > 28 kg/m included.², and glucose on admission. HDL-C < 1.03 = TRUE は、陽性事象の予測に最も大きく貢献しました。 陽性事象の予測における BMI >Contribution of 28 = FALSE was the opposite of BMI > 28 = TRUE. Obesity has also been suggested to cause disease.

interpretation of the model. We used two methods of model interpretation. (a) nomogram. The trends and magnitudes of the impact of the nine main factors on positive event predictions can be observed in the figure. Glucose on admission, BMI >28, age, NAFLD, UA, and smoking are risk factors for PPDM-A. In contrast, cardiovascular disease, ALT and HDL-C < 1.03 are negative predictors. (B.) the Sharpley value was used to describe the influence of the model on the predictions. HDL-C < 1.03、BMI > 28, admission blood glucose was the main factor affecting prediction. BMI > 28, cardiovascular disease, HDL-C < 1.03, smoking are logistic variables, where 0 is FALSE and 1 is TRUE.

Individual diagnosis

We used a RL L1 (C = 1) model built with 9 features to assess the main impact on the prediction of the 6 samples using the Sharpe value.As a result, the main contributions to the optimistic prediction for sample 1 are (BMI > 28 = 0) = FALSE, (HDL-C < 1.03 = 0) = FALSE, with contributions of 0.68 and 0.06, respectively. indicated (Fig. FourA). The probability of an optimistic prediction for sample 1 was 0.83.The main contributions to the optimistic prediction for sample 2 were (BMI > 28 = 0) = FALSE, (HDL-C < 1.03 = 1) = FALSE, with contributions of 0.62 and 0.22, respectively (Fig. FourB). The probability of predicting this sample to be positive was 0.74. For example 3, (BMI > 28 = 0) = FALSE, (HDL-C < 1.03 = 1) = FALSE は、0.65、0.21 の正の予測可能性に寄与しました。 したがって、BMI > 28 kg/m² was the major risk factor in this sample (Fig. FourC).Multiple clinical information for samples 4, 5, and 6 did not contribute much to a positive prediction (Fig. FourD, E, F). The probability of predicting the development of PPDM-A in each of these samples was less than 0.13.

Individual diagnosis. We examined risk factors for 3 positive and 3 negative predictive samples in the predictive set. (a) Patient 1 had a BMI > 28 kg/m² A predicted probability of developing diabetes of 0.83 as the main risk factor. (B.) Patient 2 had a BMI > 28 kg/m², HDL-C > 1.03 mmol/l was the main factor, with a predicted probability of developing diabetes of 0.74. (C.) Patient 3 had a BMI > 28 kg/m² A predicted probability of developing diabetes of 0.89 as the main risk factor. (D.,picture,debt) Patients 4, 5, and 6 have no significant contributors to predicting a positive event, and all have predicted probabilities of developing diabetes less than 0.13. The contribution of risk factors to this patient can be seen graphically. Red represents positive contributions and blue represents negative contributions.