Liver Cirrhosis Diagnosis Prioritizing Algorithm Based on Electronic Health Records.

Overview

The investigators use machine learning capabilities on massive electronic health records for the purpose of developing a model that prioritizes individuals at high risk of progressing to liver cirrhosis, and validating it with participants that the model found to be at high risk. constructing and validating a reliable model, with sufficient accuracy to justify further and expensive means of detection, will enable treating patients with damaged liver at an early enough stage to allow improvement of the liver condition.

In this study the investigators harness modern capabilities of machine learning in the field of hepatology for developing a model that can identify prioritize individuals at high risk of progressing to liver cirrhosis at an early and treatable stage. Cirrhosis is an advanced state of liver disease that usually manifest when the liver is already severely damaged, without many treatment options and gloomy prognosis. There are currently 2 means for diagnosis, the first is liver biopsy that is costly and inflicts pain to the patients, and has its own risks. The second is a designated imaging test, such as Fibroscan, which is safe and painless but also too expensive than can be doable as a broad screening tool. Scores that calculates higher probability for a liver disease have already been developed, but with lower predictive strength than suitable to justify further examination towards detection. The study comprises of 4 distinct phases: 1. Model development. A machine learning model predicting time-to-event for liver cirrhosis diagnosis will be developed based on Electronic Health Records. Records are anonymized and all work is performed on a designated server. 2. Anonymized Electronic Health Records latest lab test results and diagnoses from Clalit healthcare's North district will be obtained. On those records the trained model from phase 1 will run to predict time-to-event for liver cirrhosis diagnosis. Via predictions individuals will be ordered by risk. Via the deanonymized records, available only to clinicians, 20 individuals of highest risk will be observed. This includes measuring latest FIB-4 scores, viewing prior diagnoses and tests, as well as textual information from physicians. These individuals are not invited to a visit and are only viewed retrospectively through their records. 3. Upon results from phase 2 the machine learning model from phase 1 will be revised. This includes possible alterations such as revisions of inclusion/exclusion criteria, change of lab tests given as input to the model, etc. 4. As in phase 2, updated anonymized Electronic Health Records from Clalit healthcare's North district will be obtained. The updated model from phase 3 will run to predict time-to-event for liver cirrhosis diagnosis. In addition, all individuals will have their FIB-4 score computed. A ranked list of the top individuals with highest predicted risk predicted by the model from phase 3 and top individuals with the highest FIB-4 score will be constructed. Approximately a fourth of the individuals will come from the FIB-4 score group, age and gender matched to the prediction group. Group identity will remain unknown to clinicians, maintaining a double blinded study. Individuals will be invited to the clinic for checks. At the clinic individuals will undergo Fibroscan, height and weight measurements, answer the WHO Alcohol Use Disorders Identification Test (AUDIT) questionnaire. Furthermore, their files within the Electronic Health Records will be open and existing diagnoses, lab tests and medication prescriptions be collected.