AI-Created Diagnostic Scores for Emergency Department Patients with Breathing Difficulties
Ellen Tolestam Heyman
Content
Background
Half of all adult emergency department (ED) visits with a complaint of breathing difficulties, or dyspnoea, involve acute heart failure (AHF), exacerbation of chronic obstructive pulmonary disease (eCOPD), or pneumonia, which are often misdiagnosed and mistreated. We aimed to create an artificial intelligence (AI) diagnostic decision support tool to detect patients with AHF, eCOPD, and pneumonia among dyspnoeic adults at the beginning of their ED visit, before any blood tests, imaging, and physician assessment.
Methods
In this cross-sectional study, we included all ED visits of patients aged 18 years or older with dyspnoea at two regional Swedish EDs between 01/07/2017 and 31/12/2019. Gold standard diagnoses were defined as the subsequent in-hospital or ED discharge notes, and a subsample was manually reviewed by emergency medicine experts. We developed a novel neural network model and analysed unselected data from a complete regional healthcare system, along with ECGs and socioeconomic factors, by taking the event time and clinical context of the data into account. Cohort performance was compared to a simpler CatBoost model.
Results
We included 10,869 visits, with 15.1% having AHF, 13.6% eCOPD, and 13.1% pneumonia. Among the 11,656 variables, the median number of unique variables selected per ED visit was 187 (interquartile range 111–307). The model’s performance had a median micro AUROC of 87.8% (2.5–97.5 percentile; 86.4–89.3%), compared to the simpler CatBoost baseline model AUROC of 81.4% (77.5–86.6%). Aggregating the unique sets of variables
resulted in a cohort list of 2,064 diagnostic variables. Age, ECGs, previous diagnoses, and medication were considered important by the AI model, while sex, vital signs, and socioeconomic factors were deemed almost non-predictive.
Interpretation
By analysing all data, more aspects of data, and individualised variable sets through neural networks, we personalise and improve diagnostics in the emergency department. Our model
is generic and might be tested for other complaints and diagnoses.
Further information
At the bottom, the top 100 diagnostic variables of the AI-generated variable list are shown.
Researchers
Ellen Tolestam Heyman (1,2), Awais Ashfaq (3,4), Ulf Ekelund (2,5), Mattias Ohlsson (4,6), Jonas Björk (7,8), Alexander Marcel Schubert (9), Ardavan M. Khoshnood (10,11), Markus Lingman (3,4,12)
(1) Department of Emergency Medicine, Halland Hospital, Region Halland, Sweden
(2) Emergency Medicine, Department of Clinical Sciences Lund, Lund University, Lund, Sweden
(3) Halland Hospital, Region Halland, Sweden
(4) Center for Applied Intelligent Systems Research (CAISR), Halmstad University, Halmstad, Sweden
(5) Skåne University Hospital, Lund, Sweden
(6) Centre for Environmental and Climate Science, Lund University, Lund, Sweden
(7) Department of Laboratory Medicine, Division of Occupational and Environmental Medicine, Lund University, Lund, Sweden
(8) Clinical Studies Sweden, Forum South, Skåne University Hospital, Lund, Sweden
(9) Department of Computational Precision Health, University of California, Berkeley, USA, and Department of Computational Precision Health, University of California, San Francisco, USA
(10) Emergency Medicine, Department of Clinical Sciences Malmö, Lund University, Lund, Sweden
(11) Skåne University Hospital, Malmö, Sweden
(12) Department of Molecular and Clinical Medicine, Institute of Medicine, Sahlgrenska Academy, University of Gothenburg, Gothenburg, Sweden
Funding
These studies were part of the AIR Lund (Artificially Intelligent use of Registers at Lund University) research network in Lund, Sweden. The work was funded by the Swedish Research Council; the Scientific Council of Region Halland, Sweden; Sparbanksstiftelsen Varberg, Sweden; and the foundation Stiftelsen Landshövding Per Westlings Minnesfond, Sweden. The funders had no role in the study design, data collection, analysis, interpretation, writing of the report, or decision to publish.
Partners
Region Halland, Halmstad University, Lund University
Senast ändrad: