{
  "abstract": "Introduction Patients with cirrhosis are at risk of acute decompensation but robust predictors are lacking. This study developed and validated a machine learning model to predict risk of liver decompensation necessitating admission in patients being managed in a dedicated cirrhosis service.Methods Cirrhosis patients from a tertiary hepatology centre were retrospectively included. Patients were randomly divided into training and validation cohorts. Baseline clinical variables relating to demographics, disease aetiology, biochemistry, liver complications (ascites, encephalopathy, varices) and liver-related pharmacotherapy (beta-blockers, diuretics, rifaximin) were collected. Models using baseline variables to predict 6-month episodes of liver decompensation resulting in admission were developed using eight supervised learning algorithms, including the state-of-the-art generative foundation model, TabPFN, and thirteen feature selection techniques. The top three performing machine learning/feature selection combinations were combined using ensemble learning. Model performance was benchmarked against established liver prognostic scores. K-means clustering stratified patients for acute decompensation risk.Results 403 patients were included in training (n=270) and validation (n=133) cohorts. The ensemble model, consisting of logistic regression, ridge regression and neural network algorithms using baseline bilirubin, albumin, INR, ascites and diuretic use was predictive of 6-month decompensation in training (AUC 0.75, 95% CI 0.68-0.82) and validation (AUC 0.84, 95% CI 0.72-0.94). The model significantly outperformed CP grade (AUC 0.78, p<0.001), MELD (AUC 0.77, p=0.03 and UKELD (AUC 0.72, p<0.001) in predicting 6-month risk of admission due to decompensation during validation. Unsupervised machine learning identified three distinct risk groups, with significant differences in median time-to-decompensation in training (high-risk: 6.0 months, 95% CI 1.8-46.8 vs. moderate-risk: 43.9 months, 95% CI 28.3-43.9 vs. low-risk: median not reached, p<0.001) and validation (high-risk: 3.0 months, 95% CI 0.5-6.0 vs. moderate-risk: 25.7 months, 95% CI 8.8-36.7 vs. low-risk: median not reached, p<0.001) cohorts ( figure 1).Conclusions Machine learning based algorithms using routine clinical variables outperform established prognostic liver scores in predicting risk of admission due to decompensation in patients managed in a dedicated cirrhosis service. Ensemble models can stratify risk of admission, allowing for personalised care strategies including follow-up intervals and ambulatory monitoring in higher risk groups.Abstract FP29 Figure 1",
  "authors": [
    {
      "affiliations": [
        "Imperial College London, London, United Kingdom"
      ],
      "name": "Mathew Vithayathil"
    },
    {
      "affiliations": [
        "Imperial College London, London, United Kingdom"
      ],
      "name": "James Bell"
    },
    {
      "affiliations": [
        "Imperial College Healthcare NHS Trust, London, United Kingdom"
      ],
      "name": "Alexander Cole"
    },
    {
      "affiliations": [
        "Imperial College Healthcare NHS Trust, London, United Kingdom"
      ],
      "name": "Claudia Moore-Gillon"
    },
    {
      "affiliations": [
        "Imperial College London, London, United Kingdom"
      ],
      "name": "David Kockerling"
    },
    {
      "affiliations": [
        "Imperial College Healthcare NHS Trust, London, United Kingdom"
      ],
      "name": "Rooshi Nathwani"
    },
    {
      "affiliations": [
        "Imperial College London, London, United Kingdom",
        "Imperial College Healthcare NHS Trust, London, United Kingdom"
      ],
      "name": "Lucia Possamai"
    },
    {
      "affiliations": [
        "Imperial College Healthcare NHS Trust, London, United Kingdom"
      ],
      "name": "Heather Lewis"
    },
    {
      "affiliations": [
        "Imperial College Healthcare NHS Trust, London, United Kingdom"
      ],
      "name": "Nowlan Selvapatt"
    },
    {
      "affiliations": [
        "Imperial College London, London, United Kingdom"
      ],
      "name": "Maud Lemoine"
    },
    {
      "affiliations": [
        "Imperial College London, London, United Kingdom",
        "Imperial College Healthcare NHS Trust, London, United Kingdom"
      ],
      "name": "Ameet Dhar"
    }
  ],
  "title": "FP29 Machine learning and generative artificial intelligence models outperform established prognostic scoring systems in predicting acute decompensation in patients with cirrhosist",
  "uid": "a3cac702-e52e-5264-b56b-85bf71dc1bd5"
}