Healthcare Delivery Project: Identification of
Predictive Models for Treatment of Neutropenia

Summary: Every biological system can be represented by a mathematical model. The fidelity of such a model depends on the assumed form of the model, as well as the quality and quantity of data used in its determination. In this study, patients undergoing Granulocyte-Colony Stimulating Factor (G-CSF) treatment to prevent neutropenia (dangerously low neutrophil count) are viewed as systems with G-CSF as input (I) and absolute neutrophil count (ANC) as output (O).

Several treatments for cancer (radiation, chemotherapy), as well as cancer itself, can cause neutropenia. Control of neutrophil count is critical to the cancer therapy process, as neutrophils comprise a primary defense mechanism for the body. Therefore, a predictive model for ANC is very useful for treatment, for two reasons: (1) neutropenia can force the suspension of radiation treatment or chemotherapy, and a predictive model can eliminate that risk; and (2) G-CSF treatment is costly, so it should be administered as precisely (sparingly) as possible.

Given I/O behaviour over time, the goal is to fit a mathematical model to those data. Linear models were extracted from I/O data using a number of parameter estimation strategies. Specifically, two conventional on-line methods (minimization of mean square error and long range predictive identification) and two new off-line methods (optimization of a cost function with constant time delay and optimization of a cost function with a fixed pattern filter) were used. The quality of the resulting identified models was judged on the basis of the value of the root mean square estimation error and the quality of the model's predictive capability over a future horizon of interest, resulting in the following observations and conclusions:
  • There is a strong need to either increase the frequency of sample collection or to perform an in-depth investigation of the daily variation of neutrophil count to enable better data interpolation. The data available to us was irregularly and, in some cases, infrequently sampled, making it difficult to fit a good model.
  • The time delays associated with G-CSF injections seem to follow a definite pattern, which is not accommodated by standard model identification methods. The first few injections in a series seem to influence ANC more quickly than later ones.
  • Although none of the techniques produced totally faithful predictions, the performance of our new "fixed pattern filter" based on parameter optimization was significantly better in comparison with other techniques.
  • Different patients generated significantly different models. Due to the limited number of patients, the clustering of parameters with respect to age, gender and length of the treatment did not reveal any pattern.

The above study was conducted by Ms. Tamara Djokic (MScEng awarded, Spring 1999), under the supervision of Dr. Taylor (UNB), Dr. Monique Frize (Carleton University), and Dr. Mary-Frances Scully (Hemophilia Clinic Director, Faculty of Medicine, Health Sciences Centre, St. John's, NF).

Return to my Home Page

Information supplied by: Jim Taylor
Last update: 1999 October 20
Email requests for further information to: Jim Taylor (jtaylor@unb.ca)