Development And Validation Of An Artificial Intelligence Based Prognostic Prediction Model For Liver Cirrhosis

Background and ObjectivesCirrhosis is one of the leading causes of death and health burden worldwide,and its prognostic assessment is important for clinical management and treatment decision-making.Traditional prognostic assessment models,such as the Child-Pugh and MELD scores,are widely used in clinical practice but have limitations in predicting the prognosis of cirrhosis patients.In recent years,Artificial Intelligence(AI)technologies,particularly machine learning and deep learning,have shown potential in improving the accuracy of cirrhosis diagnosis,predicting the risk of complications,and assisting in the formulation of personalized treatment plans.This study aims to explore the use of AI techniques to construct a cirrhosis prognosis prediction model based on high-throughput multidimensional clinical characteristics and radiomic data.By combining clinical features,radiomic features,and deep learning algorithms,the study seeks to enhance the accuracy of survival predictions for cirrhosis patients and provide more precise decision support for clinical treatment.Part One: Construction and Validation of a Cox Regression Model for Cirrhosis Survival Prediction Based on Clinical FeaturesMethodsThis is a retrospective cohort study that included 480 cirrhosis patients hospitalized from January 2013 to December 2017 with complete records across three centers,all of whom underwent abdominal CT scans.By analyzing the patients’ baseline characteristics,laboratory test indicators,CT images,and Hepatic Venous Pressure Gradient(HVPG)values,univariate analysis and multivariate Cox regression analysis were used to identify the main factors related to cirrhosis prognosis and HVPG.Finally,we constructed a cirrhosis survival prediction model based on clinical characteristics.ResultsUnivariate analysis indicated that total bilirubin,albumin,international normalized ratio,creatinine,ascites grade,history of hepatic encephalopathy,myosteatosis,and sarcopenia were associated with the death of cirrhosis patients.The multivariate Cox regression analysis model revealed that total bilirubin(HR: 1.004,95%CI: 1.001-1.006,P=0.013),international normalized ratio(HR: 2.011,95%CI: 1.318-3.069,P=0.001),ascites grade(HR: 3.370,95%CI: 1.917-5.924,P<0.001),history of hepatic encephalopathy(HR:2.468,95%CI: 1.562-3.899,P<0.013),myosteatosis(HR: 1.543,95%CI: 1.040-2.290,P=0.031),and sarcopenia(HR: 2.341,95%CI: 1.600-3.426,P<0.001)were independent risk factors for poor cirrhosis prognosis.Among the 480 cirrhosis patients,those with both sarcopenia and myosteatosis had a 5-year mortality rate as high as 53.57%,significantly higher than other groups.L3-SMD was negatively correlated with HVPG(r=-0.266,P<0.001),while no correlation was observed between L3-SMI and HVPG.A novel prognostic prediction model was ultimately constructed,including six predictive factors such as total bilirubin,INR,history of hepatic encephalopathy,ascites grade,sarcopenia,and myosteatosis,presented in the form of a nomogram(named nomo SMCLF).The model’s Area Under the Curve(AUC)for predicting 6-month,1-year,2-year,and 5-year survival were 0.83,0.84,0.83,and 0.86,respectively,with calibration curves,time-dependent ROC analysis,and decision curve analysis validating its superior predictive accuracy compared to traditional scores.ConclusionThe novel prognostic prediction model developed by nomo SMCLF has high accuracy in predicting the survival rate of cirrhosis patients.Part Two: Application Validation of a Deep Learning Model for Diagnosing Skeletal Muscle Changes in CirrhosisMethodsThis part of the study focused on the application validation of a deep learning model for diagnosing skeletal muscle changes in cirrhosis.A publicly available code was selected and deployed for the "sarcopenia-ai" model,an AI model for the fully automatic localization and segmentation of L3 Skeletal Muscle based on Fully Convolutional Neural Networks(FCNN).The model was trained and validated using the cohort established in Part I.ResultsThe "sarcopenia-ai" model showed good consistency with manual annotations in predicting L3 vertebral level,assessing L3-SMD,and skeletal muscle area(dice coefficient of 0.95).The model achieved accuracy,precision,sensitivity,specificity,and F1 scores all over 90% in diagnosing sarcopenia and myosteatosis,demonstrating excellent diagnostic classification ability and significantly outperforming manual annotation with a time of approximately 1-2 seconds per case.ConclusionThe AI model "sarcopenia-ai" can accurately,quickly,and efficiently diagnose skeletal muscle changes in cirrhosis patients.Part Three: Establishment and Validation of a Cirrhosis Survival Prediction Model Combined with Radiomic and Clinical FeaturesMethodsThe study population was the cirrhosis patient cohort enrolled in Part I.The AI model deployed in Part II was used to automatically segment CT images and extract radiomic features.Finally,machine learning algorithms were used to select effective features and construct a novel prognostic prediction model for cirrhosis.ResultsLasso regression and XGBoost algorithms identified radiomic features related to survival prognosis.A joint model combining clinical features was constructed and presented in the form of a nomogram(nomo Rad Clin).The nomo Rad Clin model showed high predictive accuracy in both the training and validation sets,with the best discriminative ability in 6-month survival prediction for cirrhosis patients,achieving AUCs of 0.878 and0.930 in the training and validation sets,respectively.ConclusionA successful construction of a prognostic prediction model integrating radiomic features and clinical indicators using machine learning technology,named nomo Rad Clin,has been achieved.This model has an advantage in predicting the short-term survival rate of cirrhosis patients,with AUCs of 0.88 and 0.93 in the training and validation sets,respectively.Overall Conclusions1.This study found that myosteatosis is closely related to portal hypertension,while sarcopenia is not associated with it.Myosteatosis and sarcopenia are both independent risk factors for poor prognosis in patients with cirrhosis.2.The study established a novel non-invasive prognostic prediction model for cirrhosis,including TBil,INR,HE history,ascites grade,sarcopenia,and myosteatosis(nomo SMCLF model),which is superior to Child-Pugh and MELD scores in predicting the 6-month,1-year,2-year,and 5-year survival rates of patients with cirrhosis.3.The study validated a fully automatic AI model "sarcopenia-ai," which can efficiently and accurately identify the L3 vertebral level and precisely segment muscles,confirming the practicality and generalization ability of AI models in the field of medical image analysis.4.The study utilized machine learning algorithms to select radiomic features and established a prognostic prediction model combining radiomic and clinical features.This model demonstrated high diagnostic efficacy in predicting the 6-month survival of cirrhosis patients,with AUCs of 0.88 and 0.93 in the training and validation sets,respectively.