Prediction of high-risk emergency department revisits from a machine-learning algorithm: a proof-of-concept study

Study design, participants and setting

This study recruited patients who demonstrated unscheduled ED revisit within 72 hours between January 2019 and December 2021 from National Taiwan University Hospital Hsin-Chu Branch (NTUH-HCH), a tertiary centre with 829-bed capacity and more than 1700 staff. About 60 000 patients visit the ED each year; on average, 4.5% of these patients demonstrated an ED revisit after the index discharge. Patients were eligible for recruitment and analysis if they were age 20 years or older and demonstrated an ED revisit within 72 hours, whereas those who demonstrated ED revisit simply for diagnostic certificate or legal issue were immediately excluded.

Data source, features and preprocessing

For data acquisition, independent ED attending physicians retrospectively reviewed the medical charts rather than extracting information from the integrated medical database to minimise the biases and errors in the original medical record. For data dimensions, 150 features were initially included, such as age, sex, pre-existing diseases, diagnosis, final disposition and two sets of covariates from the ED index and revisit. Each set contained triage level, vital signs, chief concern, management, medication and laboratory data. Pre-existing diseases were hypertension, diabetes mellitus, coronary artery disease, cerebrovascular disease, chronic kidney disease, malignancy, chronic obstructive pulmonary disease and previous documented surgery.

Triage level was determined by the Taiwan Triage and Acuity Scale computerised triage system, which has been validated with levels 1–5 to indicate resuscitation, emergent, urgent, less urgent and non-urgent.16 Vital signs included body temperature, respiratory rate, heart rate, blood pressure and oxygen saturation.

Chief concerns, originally written on medical charts, were recorded and classified by ED attending physicians into 30 common concerns, such as headache, vertigo, chest pain, short of breath, cough, rhinorrhoea, abdominal pain, nausea, vomiting, diarrhoea, dysuria, frequent urination, retention of urine, chills, limb oedema and tube malfunction, among others.

Management included electrocardiography, chest radiography, CT, MRI, panendoscopy, colonoscopy and specialist consultation with any formal consultation from surgeons, radiologists or intensivists. Medications included analgesics and antibiotics, either orally or intravenously. Laboratory data included serum concentrations of white cell count, haemoglobin, sodium, potassium and C reactive protein; blood gas analysis; and liver function and renal function tests. Diagnosis was categorised into infection, neurological diseases, circulation diseases, respiratory disease, gastrointestinal diseases, genitourinary diseases and musculoskeletal diseases. To predict high-risk ED revisit, the features in the index visit should be included, and those in the revisit should be reasonably excluded. A total of 79 features were used for data training (online supplemental figure 1).

For data cleaning, nonsense records were first removed. For unreasonable values for the feature, we re-examined the medical record to confirm the correctness. Because the rate of missing data was 4.3%, with most missing variables missing at random, mean imputation was used to replace missing values for a specific feature by the mean of non-missing cases for that feature. For data aggregation, we aggregated the feature according to its characteristic. We set body temperature as a binary feature based on whether it ranged between 36.0℃ and 37.4 ℃ or not. In addition, blood gas features (eg, pH value, partial pressure of carbon dioxide) were also aggregated for analysis.