New advances and validation of knowledge management tools for critical care using classifier techniques

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Intensive Care Medicine

Background:

• The IDEAS for ICU's tool (version 2.0) utilized physician-selected weights for case matching in intensive care.
• Previous versions relied on subjective input for matching critical care patient data.

Purpose of the Study:

• To enhance the IDEAS for ICU's case-based reasoning (CBR) tool with quantitative matching weights and a faster engine.
• To develop and evaluate an artificial neural network (ANN) model for predicting the duration of artificial ventilation in ICU patients.

Main Methods:

• Implemented quantitatively determined matching weights and a new, faster matching engine in the CBR tool.
• Employed a back-propagation, feed-forward artificial neural network (ANN) to classify ventilation duration.
• Utilized weight-elimination techniques to reduce input variables and optimize ANN performance.
• Assessed ANN performance using correct classification rates (CCR) and Average Squared Error (ASE) with varying input variable counts.

Main Results:

• The enhanced CBR tool incorporates quantitatively derived matching weights and a more efficient matching engine.
• The ANN model successfully estimated two classes of artificial ventilation duration.
• Weight-elimination effectively reduced input variables and accelerated outcome estimation for the ANN.
• Experiments demonstrated the impact of input variable selection on ANN accuracy (CCR) and error (ASE).

Conclusions:

• Quantitative weights and improved engine speed enhance the CBR tool's utility for ICU patient data analysis.
• ANNs, optimized with weight-elimination, offer a viable method for predicting critical care outcomes like ventilation duration.
• Variable selection significantly influences ANN performance metrics in predicting patient outcomes.