Combining machine learning techniques to efficiently determine the need for HCC screening among people with HIV and hepatitis B
Description: Hepatocellular carcinoma (HCC) is a common cause of morbidity and death in persons living with HIV and hepatitis B virus (HBV) worldwide [1]. Although tenofovir-based antiretroviral therapy, the standard of care for this population, reduces the risk of developing HCC[2, 3], this risk persists even for persons without liver cirrhosis[4]. Because curative HCC treatment is possible upon early detection by liver ultrasound[5], international guidelines recommend 6-monthly screening for persons with HIV and HBV [6]. While currently used hegemonic risk prediction tools successfully capture most HCC cases in this population, their low positive-predictive value constrains their ability to prevent over-screening associated with unnecessary patient concerns, potentially harmful follow-up evaluations and high costs [7]. Reliable and individualized risk assessment tools are required to inform health professionals and patients on the need to screen for HCC and the optimal time-point to start routine screening. In order to overcome shortcomings of currently used prediction tools and measurements, I propose a novel approach combining artificial intelligence techniques to determine the need of HCC screening. I will take advantage of the comprehensive characterisation of the European cohort of patients with HIV and HBV coinfection[4, 8] with a focus on capturing the causal/sequential nature of these longitudinal data. I will apply some of the most recent advances in artificial intelligence, which were conceived and advanced in fields like spoken word and human gesture recognition [9, 10]. This study will be the first to use a combination of machine learning algorithms to explore rich longitudinal data for the development of an HCC risk prediction tool based on clinical and biological events preceding the occurrence of HCC. The overarching goal of this proposal is to develop a machine learning based tool that transforms individual data in personalized probability distributions of time to very early stage HCC. AIM 1. To identify HCC-specific patterns in patients characteristics and longitudinal laboratory measurements (“multivariate trajectories”), and to characterise the temporal relationship of these patterns with very early stage HCC.I will do this by identifying patterns among multivariate trajectories of patients diagnosed with HCC. Not all of these patterns will be predictive of HCC development, yet crucial to establish coincidences and discrepancies with patients without HCC. We will then establish, by comparison with the overall population, which of these patterns are specific to people who developed an HCC. These HCC-specific patterns will enable discrimination between people most and least likely to develop HCC. We will conclude this step with the estimation of probability distributions of time between HCC specific patterns and very early and potentially curable HCC stage, defined as a tumour of 1cm of diameter. These probability distributions will be functional to the construction of the personalized HCC risk predictions in Aim 2. AIM 2. To infer, for patients without previous evidence of HCC personalized probability distributions of time to very early HCC stage. I will infer these distributions only for patients who match HCC specific patterns. The personalized probability distributions will consist of a combination of the probability distributions (found in Aim 1) that correspond to all HCC specific patterns identified in the person’s data.AIM 3. To compare the predictive power of the proposed methodology against PAGE-B and other alternative scoring systems at the standard threshold for ultrasound screening (HCC risk of 0.2% per year). This project could establish novel methodological pipelines for screening various health conditions based on longitudinal clinical data, and identify optimized and personalized HCC screening algorithms.
Amount: CHF 87,370
Funding type: grant
Funder scheme: Spark
Start date: 2024-03-01
End date: 2025-06-30
Investigator: Salazar-Vizcaya, Luisa
Institution: University of Berne – BE
Funded outputs count: 0
Funder award ID: 221410
OpenAlex ID: G2913874131