The Artificial Intelligence Clinician learns optimal treatment strategies for sepsis in intensive care

Abstract

In this document, we explore in more detail our published work (Komorowski, Celi, Badawi, Gordon, & Faisal, 2018) for the benefit of the AI in Healthcare research community. In the above paper, we developed the AI Clinician system, which demonstrated how reinforcement learning could be used to make useful recommendations towards optimal treatment decisions from intensive care data. Since publication a number of authors have reviewed our work (e.g.

Model‐Informed Artificial Intelligence: Reinforcement Learning for Precision Dosing

“…As huge amounts of data from multiple sources become available, approaches based on machine learning will build on the knowledge obtained from mechanistic models and will be the basis to inform precision dosing. Typically, such big data will come from real‐world patient experience and will enable development of dose individualization algorithms to improve drug dosing after the initial drug approval . As we saw in the specific case of patient management in the intensive care unit, big data even offers the possibility to replace mechanism‐based models with artificial neural networks, which are, in turn, embedded into RL approaches to identify optimal dosing policies at an individual level …”

“…Typically, such big data will come from real-world patient experience and will enable development of dose individualization algorithms to improve drug dosing after the initial drug approval. 15 As we saw in the specific case of patient management in the intensive care unit, big data even offers the possibility to replace mechanism-based models with artificial neural networks, which are, in turn, embedded into RL approaches to identify optimal dosing policies at an individual level. 20 However, it is important to investigate if RL can also be applied in early development to develop precision dosing algorithms during phase I or II development that could be included in confirmatory phase III trials allowing approval of new drugs with high-quality precision dosing.…”

“…For instance, in ref. 15, the prediction of the value function was trained on 170,000 clinical records. Such large datasets cannot always be collected and made available.…”

“…Komorowski et al and Barbieri et al have recently applied such an approach for the treatment of patients with sepsis and anemia, respectively. 14,15 The computer learned the experience of many previously treated patients so that it could achieve very good results in new patients. As expected, the amount of data needed for such methods to perform is high.…”

Identifying regulatory and spatial genomic architectural elements using cell type independent machine and deep learning models

“…In the last five years machine and deep learning technologies have provided revolutionary solutions to many difficult problems in biomedicine [40]. Their utility in genomics [41], biomedical image classification [42], protein structure prediction [43] and many clinical applications [44,45,46] have been demonstrated. In this study we investigate the possibility of using combinatorial histone modification patterns, detected with machine and deep learning approaches to identify both spatial and functional chromatin interactions and architectures.…”