Ethical application of Artificial Intelligence to improve prediction of myocardial infarction in critically ill patients

Given the increasing availability of complex data streams, ICUs are optimally placed to develop and deploy data-driven interventions and innovations, including the use of Artificial Intelligence (AI). Critically unwell patients are unstable and may deteriorate quickly, requiring immediate intervention from clinicians. AI has the potential to enable early identification of deteriorating patients facilitating early intervention. However implementation faces several barriers including lack of clinical acceptance of overly simplified non "explainable" models, data management and model generalisability. Most AI models in ICU remain within the testing and prototyping environment, and no published studies to date have evaluated an AI model integrated in ICU routine clinical practice.

ICU-HEART (PI Annemarie Docherty, Wellcome £2.5million) seeks to improve survival and quality of life in critically ill patients by developing and delivering the systematic diagnosis, prediction and prevention of myocardial infarction (MI) for the first time using routinely collected multimodal high dimensional data. This project aligns with the goals of the EPSRC AI Hub for Causality in Healthcare AI with Real Data, CHAI (Sohan Seth).

There will be differing risk of MI across patients (static factors: age, comorbidity, diagnosis) and within patients over time (dynamic factors: treatment decisions, treatment response, complications). I will develop algorithms in line with TRIPOD-AI and RECORD guidelines, using multimodal data that identify patients at risk of MI firstly at ICU admission and subsequent algorithms for dynamic changes during their ICU stay. Risk prediction models will contain patient, disease and treatment variables beyond ICU admission which will be drawn from granular data within the ICU clinical information system in combination with traditional data sources. Analysis will be undertaken offline initially within the DataLoch environment, before moving to a live test bed environment.

In this complex, high-stakes setting, model utility is not solely determined by traditional metrics such as predictive accuracy. Particular attention must also be paid to human-algorithm interactions, as clinician uncertainty and model interpretability temper the value of any predictions made. I will explicitly embed consideration of these issues in the model development process, interleaving the technical concerns of data science with the fundamental critiques of philosophy.

ICU-HEART is an established interdisciplinary collaboration between Usher, Informatics and Engineering. I will take advantage of these cross-disciplinary links to develop expertise in novel causal inference techniques including G methods, multi-state modelling and Reinforcement Learning, to account for treatment affected time-varying confounding.