Name
Early detection of sepsis utilizing deep learning on electronic health record event sequences
Alternative Bezeichnung
n.a.
Kurzbeschreibung
Deep learning system for early detection of sepsis on multi-center danish dataset present outside ICUs.
Endpunkt
sepsis-positive or sepsis-negative
Input
Electronic health record (patient administration system) Diagnoses (international classification disease – 10; ICD-10), procedures (NCSP – the NOMESCO Classification of Surgical Procedures), booking information, health content (structured notes containing physiological measurements, symptom classifications, check box data such as smoking and exercise habits) Electronic health record (medication module) Dates and times for prescriptions and dispensing together with information on ingredients, dose, administration routes. Laboratory system Microbiology and blood gas analysis Medical imaging system Image descriptions from computed tomography, magnetic resonance imaging, ultrasound, X-ray, positron-emission tomography National patient register Hospital admissions, diagnoses (ICD-10), procedures (NCSP) Civil registration system Patient demographics: age, address, and marital status
Modell
Sequential CNN-LSTM model (Long-term Recurrent Convolutional Network) (Vergleichsmodelle: Gradient Boosting-Vital, Non-Sequential Multilayer perceptron)
Datenbasis
Retrospective data from multiple Danish hospitals over a seven-year period (all citizens in 4 municipalities >18y) 52,229 data-points (“contacts”) in “full” data-set 3126 data-points (“vital-sign” data-set)
Performance Metrics
Best performance: CNN-LSTM; AUROC 0.856 (3 h before sepsis onset) to AUROC 0.756 (24 h before sepsis onset)
Features
See input; must include complete vital signs = >1 event: systolic blood pressure, diastolic blood pressure, heart rate, respiratory rate, peripheral capillary oxygen saturation, and temperature
Validierung
Split: 80 (Training) 10 (Validation) 10 (Test data), five-fold cross validation, oversampling
Verfügbarkeit
Kein markterhältliches Medizinprodukt
Erstveröffentlichung
April 2020
Status
Kein markterhältliches Medizinprodukt
Quellen
https://www.sciencedirect.com/science/article/pii/S0933365719303173