Early detection of fatty liver disease in primary care

Introduction: fatty liver disease is an increasing global burden, ranging from simple steatosis todecompensated liver cirrhosis with high mortality rates. Fatty liver disease can be caused by excessivealcohol use (alcohol-related liver disease, ArLD) or by obesity, metabolic syndrome, type-2 diabetes(non-alcoholic fatty liver disease, NAFLD), or a combination of some of the aforementioned. Diagnosisof fatty liver disease can be performed with accurate precision using liver biopsies and, more recently,non-invasive tools such as transient elastography (TE). However, these tools are typically limited tospecialized in-hospital departments. Due to the high, and increasing, prevalence of patients at risk ofdeveloping liver fibrosis, the healthcare system is faced with an increasing burden from referrals ofat risk patients. To solve this, simple diagnostic tools available in primary care are highly needed forpromoting timely disease detection and optimizing referral pathways.

Aims: In the first paper,we investigated the referral patterns of patients with suspicion of liver disease from primary care to secondary care. In the second paper, we evaluated standard liver function tests used to diagnose advanced liver fibrosis in primary care. Furthermore, we evaluated how doctors perform when assessing the risk of fibrosis and cirrhosis in real-life cases. In the third paper, we developed and validated a statistical model to meet the diagnostic challenge of liver disease in primary care. In the fourth paper we developed a novel model based on artificial intelligence (AI), that enabled us to predict significant liver fibrosis in low-prevalence populations using routinely collected patient data.

Methods: In the first paper, we read all referral letters during a two-year period (2016+2017) that were submitted to the department of Gastroenterology and hepatology at Odense University Hospital. We then followed up by looking at patient records of all patients that were referred with a suspicion of liver disease in order to map the outcome of the referral for each individual patient. The assessment of the doctor’s performance was performed on a random sample of hepatologists and primary care physicians. The second paper is based on a liver-biopsied prospective cohort study of 225 patients with prior or ongoing alcohol overuse. In the third paper, we developed and validated a diagnostic algorithm called LiverPRO to detect liver fibrosis in primary care. In the fourth paper, we had the same purpose in mind as in the LiverPRO paper but wanted to advance the methodologies to explore the potential of machine-learning models. We built six ensemble-learning models (called LiverAID) with different complexities using a prospective screening cohort of 3,352 asymptomatic subjects.

Results: We found that more than half (54%) of the individuals that are referred for specialist evaluation on suspicion of liver disease are liver healthy, and referrals could potentially have been avoided saving healthcare resources. Liver healthy was defined as no presence of severe liver diseases, such as liver inflammation, fibrosis, or cirrhosis. Liver steatosis was considered liver healthy in this study. We found that 17% of the patients were referred too late, with decompensated liver cirrhosis. Furthermore, we have shown that standard blood samples are inadequate to detect liver fibrosis and that the doctors’ abilities are limited with substantial divergence in kappa values In contrast, we found that the LiverPRO score could diagnose fibrosis with a diagnostic accuracy of 0.85 and 0.80, in both a high-prevalence and a low-prevalence cohort, respectively. Combined with a very low cost and an easily interpretable score from 0-100%, the score could be a useful screening tool and decision aid for primary care referral pathways. Finally, we found that the LiverAID models appropriately identified patients with significant liver fibrosis (TE ≥8 kPa), (AUC 0.86-0.94) and had a significantly superior discriminative ability (AUC 0.70-0.76) than standard blood-based indices. 

Conclusion: The referral patterns of patients with suspicion of fatty liver disease from primary care are inadequate, and existing tools such as liver function tests are of limited guidance. Algorithms based on non-invasive markers are potentially strong solutions to solve this growing challenge in the healthcare system.