ISPOR POSTER ID 117783
Introduction
Electronic health records (EHR) are digital repositories that contain information about patients’ medical history including symptoms, clinical examination findings, test results, procedures and prescriptions (1). All the stored variables are usually related to an outcome of each patient. These outcomes can be used as labels to train machine learning (ML) algorithms and thus build automatic classifiers for particular clinical conditions (2).Once trained, automatic classifiers offer several advantages for the diagnosis of pathologies: they don't rely heavily on human assistance, they can diagnose patients anywhere in the world, they can operate tirelessly for months and years and they are capable of determining which of all the variables used are the most influential in the process (3).
Objectives
The objective of this work is to provide an overview of the general state of research in the field of ML and EHR, the level of development by therapeutic area, the techniques, methods, and procedures used to date
Methods
Our systematic review was conducted searching for articles in English from inception and up to September 8, 2021. The databases analyzed were Scopus and Google Scholar and they were screened for titles, abstracts and keywords containing the words 'machine learning' AND 'electronic health records'. The search for articles was circumscribed to only full-text articles. After the identification of the articles in each database, both pools were unified, removing the duplicates (Figure 1).
Results
Of the studies analyzed (n = 117), they belonged to the following medical specialties in order of frequency: cardiovascular (n = 27), psychiatry (n = 19), oncology (n = 14), diabetes (n = 13), neurology (n = 9), infectology (n = 8), nephrology (n = 4), rheumatology and gastroenterology (each n = 3), emergentology, hepatology, gynecology, metabolism and ophthalmology (each n = 2), and finally surgery, pneumonology, traumatology and dermatology (each n = 1).
The country that produced the highest amount of articles was the USA (n=68) followed by China (n=14), UK (n=10) and Israel (n=3). With a production of 2 articles each, Spain, Sweeden, Taiwan and Italy were found. Regarding the language of the EHRs, most of them had records in English (n=85), in Chinese (n=16) and in Spanish (n=3). The different medical specialties presented different use of ML methods, being the most frequent the random forest (RF) (Figure 2). They also presented variations in number of patients and duration (Figure 3).
Conclusions
After a meticulous analysis of the data, it can be concluded that the different medical specialties use particular techniques to analyze their EHRs. Although there are specialties such as cardiovascular and psychiatry that had a lot of development, it would also be necessary to promote the development of other less prolific ones such as ophthalmology or dermatology. Finally, more development in this area would be optimal to enhance health outcomes and the cost effectiveness of procedures in health systems.
References
Steele, A. J., Denaxas, S. C., Shah, A. D., Hemingway, H., & Luscombe, N. M. (2018). Machine learning models in electronic health records can outperform conventional survival models for predicting patient mortality in coronary artery disease. PloS one, 13(8), e0202344.
Lv, H., Yang, X., Wang, B., Wang, S., Du, X., Tan, Q., ... & Xia, Y. (2021). machine learning–driven models to predict prognostic outcomes in patients hospitalized with heart failure using electronic health records: retrospective study. Journal of medical Internet research, 23(4), e24996.
Awan, S. E., Bennamoun, M., Sohel, F., Sanfilippo, F. M., Chow, B. J., & Dwivedi, G. (2019). Feature selection and transformation by machine learning reduce variable numbers and improve prediction for heart failure readmission or death. PloS one, 14(6), e0218760.
Note: Access the link to download the full poster with figures.
Comentarios