Dichotomization of ICU length of stay based on model calibration

  • Authors:

  • Marion Verduijn;Niels Peek;Frans Voorbraak;Evert de Jonge;Bas de Mol

  • Affiliations:

  • Dept. of Medical Informatics, Academic Medical Center, University of Amsterdam (AMC UvA), Amsterdam, The Netherlands;Dept. of Medical Informatics, Academic Medical Center, University of Amsterdam (AMC UvA), Amsterdam, The Netherlands;Dept. of Medical Informatics, Academic Medical Center, University of Amsterdam (AMC UvA), Amsterdam, The Netherlands;Dept. of Intensive Care Medicine, AMC UvA, Amsterdam, The Netherlands;Dept. of Cardio-thoracic Surgery, AMC UvA, Amsterdam, The Netherlands

  • Venue:
  • AIME'05 Proceedings of the 10th conference on Artificial Intelligence in Medicine
  • Year:
  • 2005

Quantified Score

Hi-index 0.00

Visualization

Abstract

This paper presents a method to choose the threshold for dichotomization of survival outcomes in a structured fashion based on data analysis. The method is illustrated with an application to the prediction problem of the outcome length of stay at Intensive Care Unit (ICU LOS). Threshold selection is based on comparing the calibration of predictive models for dichotomized outcomes with increasing threshold values. To quantify model calibration a measure insensitive to class unbalance is used. The threshold value for which the associated predictive model has superior calibration is selected, and the corresponding model is used in practice. Using this method to select the threshold for ICU LOS, the best model calibration is found at a threshold of five days.