Host: The Japanese Pharmacological Society
Name : The 97th Annual Meeting of the Japanese Pharmacological Society
Number : 97
Location : [in Japanese]
Date : December 14, 2023 - December 16, 2023
Since the pandemic of COVID-19, various strains of the virus have appeared causing stringency on medical resources each time. Predicting the short- and long-term prognosis of severity of COVID-19 patients will not only provide indicators for prevention or treatment, but it is also extremely useful in building systems for securing medical supplies, resources, and hospital beds.
We have developed AI models that predict long- and short-term changes in disease severity (increasing severity, recovery, and death) using individuals' conditions and medical histories for the patients infected with COVID-19 and hospitalized in ICU, during the period from 2020 to 2022, so-called 1st to 6th pandemic wave. We adopted Deep Insight method which converts numerical data into an image format to build a classifier using deep learning framework, in which the accuracy and kappa scores are as high as 0.9 and 0.8, respectively towards the validation data of the specific period identical to the training data.
However, in reality, building and optimization of AI model with the latest data requires a huge amount of procedures to ensure the security of personal information for use of medical data for analysis outside the hospital, and a lot of interactions and resources in and between multiple organizations including data cleansing, AI model optimization, data and model management.
Here we propose and demonstrate a framework for Medical MLOps, using the above COVID-19 severity prediction model as an example, towards the goal of end-to-end automatic orchestration of data flow, model tuning and model deployment at clinical site for sustainable operation for medical AI models.
