Non-alcoholic Fatty Liver and Liver Fibrosis Predictive Analytics: Risk Prediction and Machine Learning Techniques for Improved Preventive Medicine

Abstract

Non-alcoholic fatty liver disease (NAFLD) is the most common liver disease worldwide, with a prevalence of 20%–30% in the general population. NAFLD is associated with increased risk of cardiovascular disease and may progress to cirrhosis with time. The purpose of this study was to predict the risks associated with NAFLD and advanced fibrosis on the Fatty Liver Index (FLI) and the ‘NAFLD fibrosis 4’ calculator (FIB-4), to enable physicians to make more optimal preventive medical decisions. A prospective cohort of apparently healthy volunteers from the Tel Aviv Medical Center Inflammation Survey (TAMCIS), admitted for their routine annual health check-up. Data from the TAMCIS database were subjected to machine learning classification models to predict individual risk after extensive data preparation that included the computation of independent variables over several time points. After incorporating the time covariates and other key variables, this technique outperformed the predictive power of current popular methods (an improvement in AUC above 0.82). New powerful factors were identified during the predictive process. The findings can be used for risk stratification and in planning future preventive strategies based on lifestyle modifications and medical treatment to reduce the disease burden. Interventions to prevent chronic disease can substantially reduce medical complications and the costs of the disease. The findings highlight the value of predictive analytic tools in health care environments. NAFLD constitutes a growing burden on the health system; thus, identification of the factors related to its incidence can make a strong contribution to preventive medicine.

Appendices

Appendices

Appendix A: Independent Variables

The independent variables were composed of demographics, athletics habits & exercise test results, metabolic components (number of components of the metabolic syndrome, waist-hip ratio, BMI, glucose and hemoglobin A1c blood tests), routine blood counts, chemistry and lipid profile, history of hypertension, dyslipidemia and cardiovascular diseases as well as medication intake. Lifestyle and subjective stress questionnaire scores were also obtained. For the dataset, blood samples were drawn upon the participants’ arrival at the center after a 12-h overnight fast. All blood tests were analyzed in the fasting state, and analyzed in routine laboratories working under the ISO 9001:2008 quality assurance stipulations. The concentrations of glycated hemoglobin (A1C) and the concentration of total hemoglobin were measured, and the ratio reported as % A1C. A1C levels were categorized into three states, healthy <5.7%, pre-diabetic 5.7–6.4% and diabetic >6.5% according to the American Diabetes Association (ADA) guidelines. Evaluation and diagnosis of metabolic syndrome (and its components) were performed with reference to the joint interim statement of the International Diabetes Federation Task Force on Epidemiology and Prevention, the National Heart, Lung, and Blood Institute, the American Heart Association, the World Heart Federation, the International Atherosclerosis Society and the International Association for the Study of Obesity [42]. Briefly, an elevated waist circumference was defined as ≥94 cm (37 in.) in men and ≥ 80 (31.5 in.) in women, as recommended for individuals of European and Middle Eastern descent. Elevated triglycerides were defined as ≥150 mg/dl (1.7 mmol/l) or on drug treatment for elevated triglycerides. Reduced high-density lipoprotein-cholesterol (HDL) was defined as <40 mg/dl (1.0 mmol/l) in men and < 50 mg/dl (1.3 mmol/l) in women. Elevated blood pressure was defined as ≥130 mmHg for systolic blood pressure or ≥ 85 mmHg for diastolic blood pressure or on antihypertensive drug treatment in a patient with a history of hypertension. Elevated fasting glucose was defined as ≥100 mg/dl (5.55 mmol/l). The diagnosis of metabolic syndrome was based on the existence of at least three abnormal findings out of the five mentioned above [43]. The exercise ECG stress test was performed according to the Bruce protocol. ECG results were manually reviewed on the spot by a cardiologist. All participants were asked whether they had received any previous diagnosis of liver or cardiovascular disease or diabetes as part of their medical history questionnaire.

Appendix B: Definition of Gains and Lift

At a specified percentile level, the lift represents the ratio of the incident frequency up to this percentile to the total incident frequency in the full dataset. The lift chart sorts the predicted probabilities [44] in descending order and shows the corresponding curve. The lift chart depicts the extent of improvement of the prediction rate returned by our model compared to random expectation. The X axis shows the percentages (e.g. 10% refers to the top decile) and the Y axis shows the value of the lift. For a model with good prediction, the lift values will start above 1 and gradually decrease to 1. If the model does not improve the prediction, the lift values are around 1. Gains at a specified percentile level indicate the ratio of the cumulative number of incidents up to this percentile to the total number of incidents in the full dataset. The gains chart shows how widely the net needs to be cast to capture a given percentage of all the incidents in the dataset. For a model with good prediction, the gains values rise steeply in the 100% direction and then continue in a straight line. If the model does not improve the prediction, the gains values gradually increase from 0 to 100%. The visualization of the predictive analytic outcomes have become increasingly important to interpreting many graphs (e.g. ROC) and more recently Lift and Gains [45, 46].

Appendix C: Decision tree for FLI-CHAID model

Fig. 7

Decision tree for FLI-CHAID model

The figure shows part of the decision tree developed on the full dataset (the entire tree is very spread out and contains many branches) to predict FLI. The dataset contained 7581 records. The first split (tree root) was made on the basis of the BMI on the last visit (BMI-last), and split into several branches. The branch we selected presents obese individuals on their last visit (BMI >28.7, n = 760). Of these, 373 (49%) had a FLI < 60 and 387 (51%) had a FLI ≥ 60. For this subpopulation, the next split of the tree was based on the annual change in triglycerides between the first and the last visit (triglyceride velocity). For individuals with an annual change of less than 6.4 (227 individuals) the next split was by gender, and for the rest (533 individuals) the next split was by weight on the first visit. The next splits continued as depicted. For those with a weight > 87.9 kg (257 individuals), the following split was made according to the last Uric Acid value. At each stage of the tree’s growth, the prevalence distribution of FLI ≥ 60 and FLI < 60 changed. Notably, at the starting point (at the root of the tree) the distribution was 15.01%. In the gender branch, there was a prediction of 94 women with FLI < 60 (because 82% had a FLI < 60 and only 18% had a FLI ≥ 60), and 133 men with FLI < 60 (because 58% had a FLI < 60 and 42% had a FLI ≥ 60). In the weight branch, there was a prediction of 276 people to weigh less than 87.9 kg with a FLI < 60 (because 51% had a FLI < 60 and 49% had a FLI ≥ 60). For those whose weight exceeded 87.9 kg, the calculation returned 174 individuals whose last Uric Acid value was ≤6.3 who had a FLI ≥ 60 (because 63% had a FLI ≥ 60), and 83 individuals whose last Uric Acid value was >6.3 who had a FLI ≥ 60 (because 85% had a FLI ≥ 60).

Appendix D: lift and gains curves for the FIB4-CHAID models

Fig. 8

Lift chart for FIB4-CHAID

Fig. 9

Gains chart for FIB4-CHAID