ICU Length-of-Stay Prediction Using DeepHit Competing Risks

The prediction model is segmented into multiple specialized components: a severity model to assess patient condition, a mortality risk model to predict death probability, and a LOS distribution model to handle asymmetrical discharge patterns. This segmentation allows each component to specialize in capturing specific aspects of the complex relationship between patient characteristics and LOS, thereby improving overall prediction accuracy without requiring a single overly complex monolithic model.

The invention introduces an additional dimension by incorporating mortality risk as a separate predictive factor alongside patient severity. By modeling the relationship between severity, mortality risk, and LOS in a multi-dimensional space rather than a simple linear relationship, the system can capture the complex non-linear interactions that exist in ICU patient data, leading to more accurate predictions.

The system extracts only the most relevant features from electronic medical records using natural language processing and machine learning techniques. By selectively extracting critical features such as patient demographics, clinical measurements, diagnosis codes, and treatment information while filtering out redundant data, the model achieves high prediction accuracy without processing the entire medical record, thus reducing computational time and resource requirements.

Different features are extracted and processed with different levels of detail based on their importance to LOS prediction. Critical features like vital signs and lab results receive more detailed processing, while less important features are processed more summarily. This local quality approach ensures that computational resources are allocated efficiently to the features that matter most for accurate prediction.

The LOS prediction model explicitly incorporates asymmetry by using different distributional assumptions for different patient outcomes. Rather than assuming a symmetrical normal distribution, the model uses skewness-adjusted distributions that reflect the reality that LOS can be heavily skewed toward shorter stays for minor cases while having a long tail for severe cases. This asymmetric modeling approach captures discharge timing preferences and institutional patterns more accurately.

The model performs preliminary classification of patients into outcome categories (e.g., discharge alive, discharge deceased, transfer) before predicting LOS. By first determining the likely discharge disposition and then applying appropriate LOS distributions for each category, the model simplifies the overall prediction task while maintaining high accuracy, avoiding the need for a single complex model to handle all scenarios simultaneously.