Machine learning-based CAGIB scores predict in-hospital mortality in patients with acute gastrointestinal bleeding

Patient characteristics

Overall, 2467 patients with liver cirrhosis and AGIB were included (Table 1). The most common etiology of liver cirrhosis was hepatitis B virus infection (n = 1255), followed by alcohol abuse (n = 365), and hepatitis C virus infection (n = 229). Means, Meldna, and Meld 3.0 scores were 7.76, 14.09, and 14.39, respectively. Among them, 1,602 patients presented hemolysis upon admission, 1,492 were diagnosed with varicose bleeding on endoscopic examination, 1,414 were treated with endoscopic treatment, 1,498 were treated with blood transfusions while hospitalised, and 139 died. Causes of death included hemorrhagic shock (n = 71), multiple organ failure (n = 45), infection (n = 9), HCC (n = 7), HE (n = 6), and volume overload (n = 1).

1,000223334 patients each were assigned to the training and validation cohort, respectively. The training cohort showed significantly higher percentages of men (70.00% vs 65.70%, P = 0.023) and ICU admissions (15.70% vs 12.60%, P = 0.028), TBIL levels (44.59 ± 71.18 vs 40.74 ± 61.90, P = 0.012), and P = 0.012), P = 0.012), 13.89 ± 5.83, p = 0.034) than the validation cohort. However, the other variables were not significantly different between the two cohorts (Table 1).

Validation of CAGIB scores in training cohorts

In an overall analysis of 1233 patients from the training cohort, the performance of CAGIB scores (AUC = 0.789) for predicting in-hospital death was statistically shown to be child patients (AUC = 0.804, P = 0.569), MELD-NA (AUC = 0.817, P = 0.234), MELD 3.0 (AUC = 0.82, P = 0.132, P = 0.82, P = 0.132) = 0.797, p = 0.774), and Augustine (AUC = 0.814, p = 0.170) scores (Fig. 1A).

In a subgroup analysis of 736 patients with varicose bleeding, 708 patients who received endoscopic treatment, and 463 patients who received only pharmacological treatment (Supplementary Table 1), CAGIB score performance to predict in-hospital death was statistically similar to that of children, Meld-NA, Meld 3.0, Auguartin score, and Auguartin score, and Auguartin scoore. 1b – d).

Validation of CAGIB scores in the validation cohort

In an overall analysis of 1234 patients from the validation cohort, performance of Cagib scores (AUC = 0.801) to predict in-hospital death was statistically indicated by the child's gaze (AUC = 0.809, P = 0.801), MELD-NA (AUC = 0.803, P = 0.967), MELD 3.0 (AUC = 0.813, P = 0.63, P = 0.63, P = 0.63, (AUC = 0.851, P = 0.05), and Augustin (AUC = 0.830, P = 0.213) scores (Fig. 2A).

In a subgroup analysis of 756 patients with varicose bleeding, 706 patients who received endoscopic treatment, and 456 patients who received pharmacological treatment alone (Supplementary Table 2), the performance of Cagib scores to predict death in hospital was statistically similar to that of parenting, Meld-NA, Meld 3.0, Auguartin score, and Auguartin score, and Auguartin, and Auguartin score. 2b – d).

ML model based on the components of CAGIB scores in training cohorts

The LS-SVMR model (AUC = 0.986) was ANN (AUC = 0.894, P < 0.001), KNN (AUC = 0.895, P < 0.001), and decision tree (AUC = 0.632, P <0.001）よりも有意に高いAUCSを持っていました。 RFモデル（AUC = 1、P = 0.0138）よりも有意に低いAUC（図3A）。 LS-SVMRモデルのYoudenのインデックスは0.842でした。グレーゾーンアプローチにより、2つのカットオフ値は0.084と0.160でした（図4A）。グレーゾーンに関与している90人（7.30％）の患者がいました。院内死亡率は、LS-SVMRスコアがそれぞれ0.084、0.084–0.160、> It was 0.38%, 2.22%, and 64.37% in patients with 0.160 (Table 2).

In a subgroup analysis of patients with varice bleeding, the high AUC in the LS-SVMR model was 0.986 (Fig. 3B). There was a Juden index of 0.887. Due to the grey zone approach, the two cutoff values were 0.080 and 0.082. Four patients involved in the grey zone were four (0.54%) (Figure 4B). Hospital mortality rates are LS-SVMR scores<0.080、0.080–0.082、および> It was 0.16%, 25%, and 19.10% in patients with 0.082 (Table 2).

In subgroup analysis of endoscopic patients, the LS-SVMR model had a higher AUC of 0.983 (Figure 3C). There was a Juden index of 0.858. Due to the grey zone approach, the two cutoff values were 0.080 and 0.081 (Fig. 4C). There were two patients (0.28%) involved in the grey zone. Hospital mortality rates are LS-SVMR scores<0.080、0.080–0.081、および> It was 0.33%, 50%, and 24.18% in patients with 0.081 (Table 2).

In subgroup analysis of patients who received pharmacological treatment without endoscopic treatment, the LS-SVMR model still had a high AUC of 0.987 (Fig. 3D))). There was a Juden index of 0.921. Due to the grey zone approach, the two cutoff values were 0.092 and 0.157 (Fig. 4D). There were 31 patients (6.70%) involved in the grey zone. Hospital mortality rates are LS-SVMR scores<0.092、0.092–0.157、および> Patients with 0.157 were 0.26%, 6.45%, and 72.34% (Table 2).

Regardless of the overall or subgroup analysis of the training cohort, the DCA of the LS-SVMR model showed consistent net profit across various threshold probabilities, and the LS-SVMR model outperformed the “no-treat” strategy and demonstrated practical utility in decision making (Supplementary Fig. 1). Regardless of the overall or subgroup analysis, the calibration curve also showed that the LS-SVMR model achieved excellent predictive performance (Supplementary Fig. 2).

The importance of each component of the LS-SVMR model was calculated and ranked (Supplementary Figure 3).

ML model based on the component of CAGIB scores of validation cohorts

Performance of the LS-SVMR model (AUC = 0.983) to predict in-hospital deaths is also ANN (AUC = 0.849, P < 0.001), KNN (AUC = 0.699, P < 0.001), decision tree (AUC = 0.599, P < 0.001), XgBoost (AUC = 0.823, P < 0.0005 (AUC = 0.823, P < 0.0005), P < 0.0001) model of the validation cohort (Fig. 5A). Hospital mortality rates are LS-SVMR scores<0.084、0.084–0.160、および> It was 0.48%, 6.82%, and 64.71% in patients with 0.160 (Table 2).

In subgroup analysis of patients with varice bleeding, the LS-SVMR model still had the highest AUC (AUC = 0.984) (Fig. 5B). Hospital mortality rates are LS-SVMR scores<0.080、0.080–0.082、および> It was 0.15%, 0%, and 20.95% in patients with 0.082 (Table 2).

In subgroup analysis of endoscopic patients, the LS-SVMR model still had the highest AUC (AUC = 0.980) (Fig. 5C). Hospital mortality rates are LS-SVMR scores<0.080、0.080–0.081、および> It was 0.16%, 0%, and 19.15% in patients with 0.081 (Table 2).

In subgroup analysis of patients who received pharmacological treatment without endoscopic treatment, the LS-SVMR model still had the highest AUC (AUC = 0.984) (Fig. 5D). Hospital mortality rates are LS-SVMR scores<0.092、0.092–0.157、および> Patients with 0.157 were 1.06%, 19.23%, and 80.77%, respectively (Table 2).

Regardless of the overall or subgroup analysis of the validation cohort, the DCA of the LS-SVMR model showed consistent net profit across various threshold probabilities, and the LS-SVMR model outperformed the “no-treat” strategy and demonstrated practical utility in decision making (Supplementary Fig. 4). Regardless of the overall or subgroup analysis, the calibration curve also showed that the LS-SVMR model achieved excellent predictive performance (Supplementary Fig. 5).