“By 2025, graph technologies will be used in 80% of data and analytics innovations, up from 10% in 2021, facilitating rapid decision making across the enterprise.”
Gartner’s Top 10 Tech Trends in Data and Analytics
This manuscript from PLOS Digital Health is a timely article on the concept of a clinician-friendly tool using artificial intelligence methodologies that are explainable.
The methodology is a massively scalable comorbidity discovery method called Poisson Binomial based Comorbidity discovery, or PBC. PBC is used to analyze a health system’s electronic health records – with over 1 million patients and over 75 million visits – with a disease network that is well suited for outcomes research but devoid of protected health information (PHI). The methodology is capable of calculating the joint distributions of multiple, conditionally dependent variables on an outcome, which is exceedingly difficult.
These explainable AI methodologies with multimorbidity networks are used to focus on cardiovascular areas of heart transplantation, sinoatrial node dysfunction, and various congenital heart diseases. This effort takes on the challenge of “ab initio discovery of comorbid clinical variables” with the complex network of demographic variables at a large scale, as well as sorting out “the intertwined impact of multiple comorbidities and demographic variables”.
After the aforementioned PBC discovery of comorbidities within the EHR corpus, Probabilistic Graphical Models (PGMs) with use of nodes (representing diagnoses, procedures, and medications) and edges (representing comorbidities with transition probabilities) are used for this outcome research and have a myriad of advantages delineated in the paper. PGM is relatively explainable compared to other methods and is well suited to this computationally intensive task.
The authors also formatted the multimorbidity networks as web-based outcomes calculators for users as an app. This latter application will be extremely useful, as clinicians often have difficulty with these types of estimations of comorbidities and outcomes.
These authors deserve much praise for having both insight and fortitude to take on a project of this scale and complexity with an innovative approach (by not using neural networks). Their PBC methodology proved superior to traditional stratification approaches for detecting comorbid relationships. With efforts such as this investigation, we see real world data and experiences using a graph approach becoming as important as traditional randomized controlled trials for clinical insights and outcomes research, especially for complex multimorbid outcomes analyses with conditional dependencies of variables.
Read the full paper here.
The post Explainable AI for predicting cardiovascular outcomes using EHRs appeared first on AIMed.
Author: aimed_aj