Deep Learning for Liver Fibrosis Triage in MASLD Using Longitudinal Electronic Health Records (NIMIT-AI)

NIMIT-AI: Neural Inference for Metabolic-liver Integrated Trajectories: Leveraging Deep Learning to Enhance Reliability in MASLD Triage

This study looks at a new computer program called NIMIT-AI (Neural Inference for Metabolic-liver Integrated Trajectories, Artificial Intelligence) that helps doctors find liver scarring early in patients with fatty liver disease.

Fatty liver disease, also called metabolic dysfunction-associated steatotic liver disease (MASLD), is a common condition where fat builds up in the liver. Over time, this can cause scarring (fibrosis). Finding scarring early helps doctors treat it before it gets worse.

Right now, doctors use a blood test score called FIB-4 to check for scarring. But this score misses many patients and cannot be calculated when blood test results are incomplete.

NIMIT-AI works differently. It reads a patient's blood test results over multiple visits, not just one visit, to spot patterns that suggest liver scarring. It was tested on 969 patients seen at Siriraj Hospital in Bangkok, Thailand between 2018 and 2022.

In testing, NIMIT-AI found liver scarring more accurately than FIB-4. It also worked even when some blood test results were missing, which happens often in real clinics.

This study did not ask patients to do anything extra. It used health records that were already collected as part of regular care.

Study Overview

• Status: Completed
• Conditions: MASLD (Metabolic Dysfunction-Associated Steatotic Liver Disease)
• Intervention / Treatment: Diagnostic test: Longitudinal electronic health record analysis

• Study Type: Observational
• Enrollment (Actual): 1351

Contacts and Locations

Study Locations

• Thailand
  • Bangkok
    • Bangkok Noi, Bangkok, Thailand, 10700
      • Faculty of Medicine Siriraj Hospital

Participation Criteria

Eligibility Criteria

• Ages Eligible for Study: Adult; Older Adult
• Accepts Healthy Volunteers: N/A

• Sampling Method: Non-Probability Sample

Study Population

Adults with confirmed metabolic dysfunction-associated steatotic liver disease (MASLD) receiving outpatient hepatology care at Siriraj Hospital, a 2,500-bed tertiary academic medical centre in Bangkok, Thailand. The population reflects a high metabolic comorbidity burden typical of urban Thai patients, with elevated rates of type 2 diabetes, obesity, and cardiometabolic multimorbidity.

Description

Inclusion Criteria:

• Age ≥18 years at index visit
• Confirmed MASLD diagnosis per Delphi consensus criteria
• At least one outpatient visit with concurrent laboratory data and FibroScan liver stiffness measurement within observation window (2018-2022)
• Receiving care at Division of Gastroenterology, Faculty of Medicine Siriraj Hospital, Mahidol University

Exclusion Criteria:

• Alternative chronic liver disease aetiology (autoimmune hepatitis, primary biliary cholangitis, primary sclerosing cholangitis, Wilson's disease, haemochromatosis)
• Chronic viral hepatitis (hepatitis B or C surface antigen positivity)
• Prior liver transplantation
• Active extrahepatic malignancy at baseline
• Insufficient longitudinal data for outcome ascertainment

Study Plan

How is the study designed?

Number of groups / cohorts

2

Cohorts and Interventions

Group / Cohort	Intervention / Treatment
Primary longitudinal cohort (≥2 visits)	Diagnostic test: Longitudinal electronic health record analysis; NIMIT-AI, a gated recurrent unit deep learning model, analyzed serial outpatient laboratory results from electronic health records collected over a 5-year observation window (2018-2022) at Siriraj Hospital. The model processed up to 10 sequential visits per patient using 18 clinical features including liver enzymes, metabolic markers, comorbidity flags, and medication exposures to predict liver fibrosis stage without requiring elastography.
Singleton sensitivity analysis cohort (1 visit)	Diagnostic test: Longitudinal electronic health record analysis; NIMIT-AI, a gated recurrent unit deep learning model, analyzed serial outpatient laboratory results from electronic health records collected over a 5-year observation window (2018-2022) at Siriraj Hospital. The model processed up to 10 sequential visits per patient using 18 clinical features including liver enzymes, metabolic markers, comorbidity flags, and medication exposures to predict liver fibrosis stage without requiring elastography.

What is the study measuring?

Primary Outcome Measures

Outcome Measure	Time Frame
Area under the receiver operating characteristic curve (AUROC) for significant fibrosis (F≥2) identification	Assessed at end of observation period (December 2022)

Secondary Outcome Measures

Outcome Measure	Time Frame
Sensitivity-constrained positive predictive value (PPV) for significant fibrosis (F≥2) at optimised classification threshold	Assessed at end of observation period (December 2022)
Diagnostic performance for compensated advanced chronic liver disease (F3-F4 cACLD) reported as one-vs-rest AUROC	Assessed at end of observation period (December 2022)
Net reclassification improvement (NRI) of NIMIT-AI versus FIB-4 at guideline-recommended threshold (1.30)	Assessed at end of observation period (December 2022)
Integrated discrimination improvement (IDI) of NIMIT-AI versus FIB-4	Assessed at end of observation period (December 2022)
Attention weight distribution across visit positions for temporal interpretability of NIMIT-AI predictions	Assessed at end of observation period (December 2022)
SHAP (SHapley Additive exPlanations) feature importance values for global model interpretability across fibrosis classes	Assessed at end of observation period (December 2022)

Collaborators and Investigators

• Sponsor: Siriraj Hospital

Study record dates

Study Major Dates

• Study Start (Actual): January 1, 2018
• Primary Completion (Actual): December 31, 2022
• Study Completion (Actual): June 16, 2024

Study Registration Dates

• First Submitted: June 18, 2026
• First Submitted That Met QC Criteria: June 26, 2026
• First Posted (Actual): June 30, 2026

Study Record Updates

• Last Update Posted (Actual): June 30, 2026
• Last Update Submitted That Met QC Criteria: June 26, 2026
• Last Verified: June 1, 2026

More Information

Terms related to this study

• Keywords: Deep Learning; Liver Fibrosis; MASLD; Gated Recurrent Unit; Non-invasive Triage; Longitudinal EHR
• Additional Relevant MeSH Terms: Pathologic Processes; Digestive System Diseases; Liver Diseases; Fibrosis; Pathological Conditions, Signs and Symptoms; Liver Cirrhosis
• Other Study ID Numbers: 338/2569(SIRB1)

Drug and device information, study documents

• Studies a U.S. FDA-regulated drug product: No
• Studies a U.S. FDA-regulated device product: No