Interested in learning what's next for the gaming industry? Join gaming executives to discuss emerging parts of the industry this October at GamesBeat Summit Next. Learn more.
Acute myocardial infarction (AMI) — or coronary heart disease — is the leading cause of death in the U.S., and by 2035, it’s estimated that nearly half of adults will suffer from some form of it. Troublingly, most incidences of AMI occur absent obvious symptoms like chest pain or shortness of breath. But researchers at Florida State University and the University of Florida, Gainesville are recruiting artificial intelligence (AI) to help predict one-year mortality in intensive care unit patients who have experienced an episode.
One-year mortality was selected as the prediction window because it would allow for comparison to other studies, the researchers wrote, and because it would take into account patients that had multiple AMI-related ICU admissions within a two-year period.
“Compared with risk assessment guidelines that require manual calculation of scores, machine learning based prediction for disease outcomes such as mortality can be utilized to save time and improve prediction accuracy,” they wrote in a paper (“Building Computational Models to Predict One-Year Mortality in ICU Patients with Acute Myocardial Infarction and Post Myocardial Infarction Syndrome“) published on the preprint server Arxiv.org. “This study built and evaluated various machine learning models to predict one-year mortality in patients diagnosed with acute myocardial infarction or post-myocardial infarction syndrome.”
To assemble a dataset, the paper’s authors sourced MIMIC-III, a freely accessible critical care database maintained by the MIT Lab for Computational Physiology containing 58,000 hospital admissions from 40,000 patients. They whittled the list down to 5,037 subjects (accounting for 7,590 admissions) by selecting features “proven to be predictors of mortality,” such as kidney and liver function, admission, demographic, treatment, lab values evaluating long- and short-term overall health, and various cardiac markers.
In the end, the team decided to look at the data based on admissions — 5,436 with a diagnosis of AMI or premenstrual syndrome (PMS), the latter of which is associated with heart palpitations — rather than individuals. That was because in some cases, patients survived a year in one admission but didn’t survive a year in another.
The researchers preprocessed the records to remove duplicates, multiple treatments for the same admission, data entry errors, and outliers. To compare performance across several different machine learning models, they tapped Waikato Environment for Knowledge Analysis (WEKA), a Java-based software developed at the University of Waikato, New Zealand.
Using Google’s open source TensorFlow framework on a PC with a 2.2GHz Intel Core i7 processor, the team trained more than a dozen classification algorithms on the corpus, including AdaBoost, Attribute Selected Classifier, Bayes Net, Classification Via Regression, and Decision Stump, to name a few.
In tests, two AI models — the Logistic Model Trees (LMT) and Simple Logistic algorithms — performed better than the rest, achieving 85.12 percent accuracy in identifying the 30 percent of patients from the dataset (1,629) who died within one year of admission. (A third algorithm, J48, followed close behind with 84.88 percent accuracy.) Interestingly, a deep neural network model — a model with layers of mathematical functions that loosely mimic the behavior neurons in the human brain — outperformed all of the machine learning algorithms studied in its ability to identify patients who died within one year.
“This reflects a common understanding in data science that there is no universally applicable algorithm that outperforms all the other algorithms all the time,” the researchers wrote. “There are many factors that can affect mortality rates following a myocardial infarction. Figuring out a way to utilize the information regarding these factors will assist in accurately predicting possible outcomes.”
The paper’s authors note that the imbalanced dataset (30 percent of one-year mortality cases) was a limiting factor for the study, as were data gaps like missing lab and chart values. But they contend that the results show that correct diagnosis and treatment of AMI have a demonstrable effect on one-year mortality.
“As seen from this dataset, there is not one specific factor that provides the needed predictability information, while being able to include all relevant criteria leads to improved predictions,” they wrote. “The improved predictability obtained by using machine learning can help at-risk patients strive for compliance to treatment plans to improve mortality risk.”
VentureBeat's mission is to be a digital town square for technical decision-makers to gain knowledge about transformative enterprise technology and transact. Learn more about membership.