Soren Sabet Sarvestany 1, Orlando Cerocchi 2, Fuad Ahmed Ali 3, Bettina Hansen, Harry Janssen, Jeff Kwong, Giada Sebastiani 4, Keyur Patel 1, Anna Goldenberg 5,6,7* and Mamatha Bhat 8,9,10*
1 University of Toronto
2 Toronto Centre for Liver Disease, University Health Network,
3 Multi Organ Transplant Program, University Health Network,
4 Mcgill University Health Centre,
5 Sick Kids Research Institute,
6 Vector Institute,
7 Computer Science, University of Toronto,
8 University Health Network,
9 Toronto General Hospital,
10 Division of Gastroenterology and Hepatology, University Health Network, University of Toronto
*Contributed equally
Patients with cirrhosis often present to medical attention once they have developed complications of end-stage liver disease. Systematic detection of subclinical cirrhosis in the general population has previously not been feasible in due to a lack of specific and accurate, and cost effective biomarkers. The goal of this study was to develop a machine learning algorithm that could detect patients with compensated cirrhosis using blood test results.
The available data consisted of F0, F1, and F4 liver biopsies on patients seen at the Toronto Liver Clinic, with associated laboratory and clinical parameters. A +/- 90 day interval was used to link blood tests to biopsies, and records missing 3 or more parameters were excluded. This data was used to train 6 classification algorithms (Support Vector Machine, Random Forest, Gradient Boosting, Logistic Regression, K-Nearest Neighbours, Artificial Neural Network, and Ensemble) with classification performance measured by sensitivity, specificity, and AUROC. The machine learning algorithm’s performance was also compared to APRI and FIB4’s performance on the same set of patients.
On a held-out test set consisting of biopsy records from 2012 to 2014, the ensemble algorithm achieved the best overall performance with sensitivity 97.2%, specificity 54.90%, and AUROC 0.859, while only being indeterminate for 17.9% of patients. It outperformed both APRI, whose sensitivity was 93.3% but was indeterminate on 49.0% of patients, and FIB4, whose sensitivity was 69.8% but was indeterminate for 31.1% of patients. On another test set of 50 patients of varying etiologies, the ensemble algorithm had a significantly higher sensitivity than a panel of 5 experts (92.3% vs. 40% respectively)
These results suggest that machine learning algorithms significantly outperform existing methods for detecting cirrhosis from commonly available blood test data, and could be a viable method for screening the general population.