JMIR Publications

ABSTRACT

Background:

Electronic Health Records (EHRs) contain information on many risk- and preventive factors and early signs, which can be used for prevention of diseases. However, the accumulating volume and variety of EHRs over time, makes it impossible for clinicians to analyse these large-scale EHRs during a medical visit. Machine learning can assist in early diagnostic tasks because it is able to handle multidimensional data and to incorporate interactions. Although machine learning have become well developed, studies mainly focus on methodologies and techniques but lack a healthcare focus.

Objective:

The objective of this scoping review is to provide an overview of the extent of evidence in relation to early detection for timely prevention of diseases, by using machine learning on longitudinal electronic health records. By reviewing machine learning attempts for a variety of diseases, generated knowledge and clinical benefits are investigated.

Methods:

A literature search was performed in collaboration with a medical information specialist in following databases: PubMed, Embase.com, Clarivate Analytics/Web of Science Core Collection and the Institute of Electrical and Electronics Engineers [IEEE] Xplore Digital Library and DBLP. Studies were eligible when longitudinal EHRs were used, aiming for the early detection of diseases by machine learning, in a prevention context. Studies with a technical focus or using imaging or hospital admission data were beyond the scope of this review. Study screening, selection and data extraction was performed independently by two researchers.

Results:

Twenty studies were included in this review and detected a variety of diseases. By developing and comparing machine learning models, knowledge about diagnostic performance, timing of detection, important predictors and clinical usefulness was created. Applying these machine learning models in practice, patients might benefit from a personalised healthcare, preliminary offside screening and ultimately prevention, with practical benefits like workload reduction and policy insights. EHR variables can successfully be used in a time-including neural network or LSTM technique.

Conclusions:

While machine learning models based on textual EHRs are still in the developmental stage, diagnostic insights can assist clinicians in an accurate and earlier detection of diseases. With the addition of personal responsible factors, a personalised healthcare and therefore targeted prevention methods can be pursued. If machine learning models become mature enough and applied at higher level, the preventive healthcare may be improved by preliminary screenings, a reduced workload and policy insights. Clinical Trial: project osf.io/8g7ep