Improving the detection of clinically significant steatotic liver disease using a machine learning algorithm in a real-world primary care population

Résumé fourni par la source

Abstract Background and aims Population screening for liver disease in high-risk groups is recommended. Community diagnosis of liver disease is a challenge due to the asymptomatic nature of disease until very advanced stages. Moreover, regional variation in testing availability can result in people with clinically significant liver disease being missed. Machine learning (ML) has been proposed as a method to reduce diagnostic error and automate screening. We present a novel machine learning derived algorithm (ID LIVER-ML) designed to predict the risk of clinically significant liver disease in a high-risk community population to identify those needing further investigations or specialist referral. Methods Using data from 2039 patients recruited to two UK cohorts, we created a parsimonious model using investigations that would be available in primary care using liver stiffness measurement as reference standard. The performance of ID LIVER-ML was compared against FIB-4 score in a second unseen hold out cohort (n=327). Results ID LIVER-ML performed well at identifying patients at risk of clinically significant liver fibrosis (sensitivity 0.90, Specificity 0.43, PPV 0.54, NPV 0.86, AUC 0.83) and outperformed conventional risk scoring systems (FIB-4: AUC 0.65; NAFLD Fibrosis Score: AUC 0.66; APRI: AUC 0.53; BARD: AUC 0.58). Conclusion Machine learning derived algorithms can help screen high risk populations in a community setting for liver fibrosis. ClinicalTrials.gov ID: NCT04666402 Impact and Implications The prevalence of steatotic liver disease is rising globally and is an increasingly significant challenge for healthcare systems. Existing risk stratification scores are not validated in a real-world cohort where patients have risk factors for multiple aetiologies of liver disease. Our work shows that a machine learning model can predict the risk of clinically significant liver disease using routine primary care data, better than existing non-invasive risk stratification tools in a real-world cohort. This highlights a potential role for machine learning in the automation of fibrosis risk assessment in primary care. Highlights - Machine learning derived algorithms can predict the risk of clinically significant liver disease in an at risk community population with a mixed aetiology of liver diseases. - The performance of the ML algorithm (ID LIVER-ML) is not affected by metabolic, alcohol, or mixed aetiologies. - ID LIVER-ML outperforms traditional risk stratification scoring systems such as FIB-4 and NAFLD fibrosis scores. - Compared to the FIB-4 score, the use of Machine Learning can reduce the need for secondary care investigations by 59%.