Document resource
Background Liver cancer is the fastest rising cause of cancer related mortality in the United Kingdom. Hepatocellular carcinoma (HCC), the most common primary liver cancer, has risen in incidence by 50% in the past decade. 1 Despite well-established guidance for HCC surveillance (HCCS) for high-risk individuals, only around 20% of HCC cases are identified at an early stage and treated with curative intent.2 Recognising that HCCS is currently suboptimal, NHS England have established minimum standards for identifying high risk patients, including the mandatory creation of a patient registry.3 EndoMinerAI is an automated disease cohort extraction tool which uses deep learning-based natural language processing (NLP) to automate the identification of patients requiring HCCS. It combines a multi-label BERT-based classifier to extract liver disease aetiologies from unstructured clinical letters with a rule-based engine that applies national surveillance criteria.Methods To assess the accuracy of EndoMinerAI in real-world clinical practice, we conducted a retrospective validation of 122 patients identified by the algorithm as eligible for HCCS. EndoMinerAI processed free-text clinical letters extracted from electronic hospital records between 2013–2023. Case notes were reviewed for each patient to determine if they met the criteria for HCCS, based on serological, radiological, histological and/or non-invasive assessments of liver disease. The last available encounter was reviewed to establish whether patients remained under active follow-up and HCCS, were lost to follow-up, or had been discharged.Results Clinical validation suggested that 100% of patients identified by EndoMinerAI were eligible for HCCS. 86% of these had objective evidence to support HCCS. Interestingly 14% did not, but the algorithm was able to identify these patients based on clinician subjective opinions in clinical letters. The most common liver disease aetiologies were alcohol-related liver disease (ARLD) (29%), metabolic dysfunction-associated liver disease (MASLD) (29%) and hepatitis B with significant fibrosis (F2–4) (16%). From this cohort, 63% of patient remain on regular HCCS with 6 monthly ultrasound scans; 13% have been lost to follow-up; 10% of patients were discharged due to protracted non-attendance.Conclusions EndoMinerAI presents a novel and robust method of identifying patients at risk of HCC from single data source clinical letters, enabling the creation of a patient registry. This automated tool can find those who have been lost to follow up and presents an opportunity to re-engage and improve adherence with HCC surveillance.References ‘Cancer Research UK,’ [Online]. Available: https://www.cancerresearchuk.org/health-professional/cancer-statistics/statistics-by-cancer-type/liver-cancer.‘NHS England,’ [Online]. Available: https://www.england.nhs.uk/long-read/hepatocellular-carcinoma-surveillance-minimum-standards/.M. e. a. Qurashi, ‘Improving hepatocellular carcinoma surveillance in the United Kingdom: challenges and solutions,’ Lancet Regional Health – Europe, vol. 43, no. 100963, 2024.