Key result
FT-transformer deep learning outperforms conventional models in predicting 28-day HF readmission with ~0.74 AUC.
Why the study?
Over 50% of heart failure patients are readmitted, worsening quality of life, making readmission prediction important to prevent unnecessary hospitalizations.
Does an FT-transformer deep learning model improve readmission prediction compared to conventional machine learning models in heart failure patients?
Observational (n=615)
Does an FT-transformer deep learning model improve readmission prediction compared to conventional machine learning models in heart failure patients?
Effect estimate: AUC 0.7434
A deep learning FT-transformer model can predict heart failure readmissions with moderate accuracy up to 6 months, outperforming conventional machine learning models.
May support FT-transformer modeling for HF readmission risk stratification; leaves open external validation and clinical utility.
Although heart failure (HF) diagnosis and treatment techniques have advanced, more than 50% of HF patients are readmitted. Readmission worsens the life quality of patients due to economic and psychological burdens. Therefore, readmission prediction for patients is important to prevent unnecessary readmissions. We used a feature tokenizer transformer (FT-transformer) to predict readmission by embedding all features and analyzing via transformer encoder. Our experiment with 615 HF patients outperformed conventional machine learning models, achieving an area under the curve of 0.7434 within 28 days, 0.7063 within 3 months, and 0.7039 within 6 months. FT-transformer can potentially improve patient outcomes by enabling early interventions to prevent readmissions.
No takes yet. Share an insight, caveat, or question.
Kim et al. (2023) conducted an observational in Heart failure (n=615). Feature tokenizer transformer (FT-transformer) vs. Conventional machine learning models was evaluated on Readmission prediction (AUC 0.7434). A feature tokenizer transformer (FT-transformer) deep learning model outperformed conventional machine learning models in predicting heart failure readmission, achieving an AUC of 0.7434 at 28 days.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: