Incident-induced attention-based deep learning model for early warning of sepsis onset

Abstract

Background

Accurate early warning of sepsis onset is crucial for reducing mortality. However, the inter-individual heterogeneity in clinical manifestations of sepsis leads to significant sparsity of data. The current time series analysis methods attempt to interpolate highly sparse sepsis data, yielding unsatisfactory results. In this study, we aimed to develop an efficient artificial intelligence approach for early warning of sepsis onset.

Methods

The I2former model, an incident-induced attention-based architecture, was proposed to address the challenges posed by sparse medical data. This model employs a novel increment entropy encoding strategy to extract clinically significant features from sparse data, effectively transforming the unavailable data into valuable insights. The training data were sourced from MIMIC-IV v2.2 and eICU v2.0, with external validation from Beijing Tsinghua Changgung Hospital. Five advanced models, including the Autoformer, Timesnet, Informer, Reformer, and DLinear, currently in use were used for comparison.

Results

Five metrics used for classification indicated that the I2former significantly outperformed the 5 advanced time series analysis methods, achieving area under the receiver operating characteristic (AUROC), area under the precision-recall curve (AUPRC), Matthews correlation coefficient (MCC), F1-score, and accuracy of 0.886, 0.529, 0.449, and 0.917, respectively. Furthermore, external validation using the data from Beijing Tsinghua Changgung Hospital demonstrated that the model provides accurate early warnings, on average of 15.5 h prior to sepsis onset.

Conclusion

Therefore, I2former is proposed for accurate early warning of sepsis onset. Five crucial metrics for classification underscored the substantial advantages of I2former in managing sparse data, while highlighting its potential application and value in the field of medical data analysis.