The implementation of health information technology aims to improve the efficiency and safety of health care. Part of this process has involved the digitalisation of patient health care records that can consequently be more easily shared across networks. These electronic health records (EHRs) can include a range of data, including demographics, medical history, allergies, test results and radiology images. The wealth of data acquired in EHRs not only provides information on individuals but as a collective also provides population-based data. Assaf Gottlieb from Stanford University, USA, Roded Sharan from Tel Aviv University, Israel, and colleagues assessed the ability of leveraging this resource in the context of clinical decision-making, with the goal of improving individual patient discharge diagnoses by integrating population-based information. Gottlieb and Sharan discuss the importance and potential impact of their findings, published in a recent study in BMC Medicine.
What are clinical decision support systems and electronic health records?
Clinical decision support systems are computer programs designed to assist physicians and other health professionals with decision-making tasks. A conceptual definition by Robert Hayward of the Centre for Health Evidence, Canada, states that these systems “link health observations with health knowledge to influence health choices by clinicians for improved health care”.
Electronic health records (EHRs) are the collection of health information about individual patients in digital form.
Why is it important to use electronic health records to help predict discharge of patients?
Clinical decision support systems typically focus on a single patient and apply manually or automatically constructed decision rules to infer a diagnosis. A large corpus of electronic health records offers a population-based viewpoint, accounting for population heterogeneity and identifying rare cases that may be missed by manually constructed clinical decision rules.
Why was it necessary to develop a new method to evaluate patient discharge diagnosis?
While a few methods had been released for predicting specific patient outcomes using large cohorts of patients, each of those was tailored to a specific condition (e.g. heart failure). It was therefore necessary to develop a general method that is applicable to (almost) every patient. The method’s predictions are based solely on similarity to a cohort of previously hospitalized patients. By this, we hope to aid the health practitioner in creating diagnoses with higher accuracy and bring to their attention similar past cases.
Can you describe the method you used to exploit electronic health records for predicting patient discharge diagnosis?
We used basic patient-specific information gathered upon admission, including medical history, blood tests, electrocardiography (ECG) results and demographics, to measure similarity among patients. For a given patient, the past diagnoses of similar patients were used to infer the discharge diagnosis.
What are the main findings of your study? What most excited/interested you about them?
The main finding of this study is that a large corpus of patient data can be exploited to predict the likely discharge diagnoses for a new patient, even with a minimal set of patient information, including medical history, blood tests performed upon admission and demographics. As our method can be readily extended to incorporate the use of additional admission information, we envision that in the future our method may incorporate medical images and patient genomic information (e.g. gene expression or genomic variation data) to produce ever more accurate predictions.
How will the results of this research impact clinical practice?
Our results may affect clinical practice in two ways: (1) aiding the clinical practitioner by supporting, or complementing, their suggested diagnoses; and (2) inferring additional patient outcomes (e.g. disease prognosis).
What are your future research directions on the use of electronic health records?
We aim to extend our method in two ways: (1) incorporate additional patient information to provide more accurate predictions; and (2) use patient similarities to additionally predict patient outcome (e.g. disease prognosis) and identify similar patients for clinical trials.
Questions from Ursula D’Souza, Senior Editor for BMC Medicine.