AI-based fingerprint index of visceral adipose tissue for the prediction of bowel damage in patients with Crohn's disease

Abstract

The fingerprint features of visceral adipose tissue (VAT) are intricately linked to bowel damage (BD) in patients with Crohn's disease (CD). We aimed to develop a VAT fingerprint index (VAT-FI) using radiomics and deep learning features extracted from computed tomography (CT) images of 1,135 CD patients across six hospitals (training cohort, n = 600; testing cohort, n = 535) for predicting BD, and to compare it with a subcutaneous adipose tissue (SAT)-FI. VAT-FI exhibited greater predictive accuracy than SAT-FI in both training (area under the receiver operating characteristic curve [AUC] = 0.822 vs. AUC = 0.745, p = 0.019) and testing (AUC = 0.791 vs. AUC = 0.687, p = 0.019) cohorts. Multivariate logistic regression analysis highlighted VAT-FI as the sole significant predictor (training cohort: hazard ratio [HR] = 1.684, p = 0.012; testing cohort: HR = 2.649, p < 0.001). Through Shapley additive explanation (SHAP) analysis, we further quantitatively elucidated the predictive relationship between VAT-FI and BD, highlighting potential connections such as Radio479 (wavelet-HLH-first-order standard deviation)-Frequency loose stools-BD severity. VAT-FI offers an accurate means for characterizing BD, minimizing the need for extensive clinical data.

Keywords: Artificial Intelligence; Health sciences.

Graphical abstract

Figure 1

Predictive performance of VAT-FI in training and testing cohorts (A–C) Plots show the receiver operating characteristic (ROC) curve of VAT-FI in the training cohort (A) and testing cohorts (B). Plot (C) shows the calibration curves for the VAT-FI in the training cohort and the total testing cohort. VAT, visceral adipose tissue; AUC, area under the receiver operating characteristic curve; FI, fingerprint index.

Figure 2

The SHAP plots and raincloud plots of VAT-FI (A) The SHAP plot illustrates the most influential features on prediction and the distribution of each feature’s impact on model, which includes a series of plots where each dot corresponds to an individual. The colors represent feature values for numeric features: red for larger values and blue for smaller. The line’s thickness, composed of individual dots, reflected the count of examples for specific values. A negative SHAP value (extending to the left) indicates reduced risk of severe BD, while a positive value (extending to the right) indicates increased risk of severe BD. (B) The weights of variables importance. The graph displays mean absolute value of the SHAP values for features in VAT-FI. (C and D) The raincloud plots show the distribution of radiomics features (C) and deep learning features (D) between groups with (red color) and without (blue color) severe BD. (E and F) SHAP individual force plots for two representative CD patients with severe (E; a 23-year-old female patient) or nonsevere BD (F; a 21-year-old male patient). The results show the impact of risk factor for prediction. Red features (left) indicate factors raising BD risk, and blue ones denote risk-reducing factors. Arrow lengths (red and blue) reflect SHAP values for each specific prediction feature. The heatmap images highlight their discrepancy in VAT between two patients with and without severe BD, whereas SAT differences were minimal. Error bars represent the 95% confidence intervals. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001. Radio293: wavelet-LHH-first-order standard deviation; Radio479: wavelet-HLH-first-order standard deviation; Radio944: log-sigma-3-0-mm-3D-first-order standard deviation. VAT, visceral adipose tissue; SAT, subcutaneous adipose tissue; FI, fingerprint index; Radio, radiomics feature; DL, deep learning feature; BD, bowel damage; SHAP, Shapley additive explanation.

Figure 3

The chord diagram and Sankey diagram of VAT-FI (A) Chord diagram illustrates pairwise correlations among top 20 fingerprint features of VAT and DSI items, with significant feature pairs and their coefficients shown below the chord graph. The red bars indicate positive correlations, while the blue ones represent negative correlations. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001. (B) Sankey diagram illustrates the putative pathways connecting “fingerprint features of VAT” with “DSI items” and ultimately influencing “BD severity.” The left nodes represent deep learning and radiomics features, middle nodes represent DSI items, and right nodes indicate BD severity. For (A) and (B), red lines indicate positive correlation, while blue lines indicate negative correlation. The thickness of lines presents strength of correlation, with thicker lines indicating a stronger correlation. (CRP, C-reactive protein; DSI, disease severity index; VAT, visceral adipose tissue; BD, bowel damage; DL, deep learning feature; Radio, radiomics feature.).

Figure 4

Comparison of the predictive performance between VAT-FI and SAT-FI (A–D) ROC curves of FIs in training cohort (A) and total testing cohort (B). Decision curves of FIs in training cohort (C) and total testing cohort (D). AUC, area under the receiver operating characteristic curve; FI, fingerprint index; ROC, receiver operating characteristic; SAT, subcutaneous adipose tissue; VAT, visceral adipose tissue.