A Fuzzy Discrete Event Systems Approach to Selecting Second-Round Combination Antiretroviral Therapy for HIV/AIDS Patients

We have recently pioneered the development of an innovative general-purpose decision-making and optimization technology, called fuzzy discrete event systems (FDES). In the previous papers, we reported results of applying FDES to selecting optimal first-round regimens for HIV/AIDS patients. In the present paper, we describe our further effort to apply the FDES framework to the second-round treatment, which is more challenging primarily due to drug resistance that occurs during the first-round treatment. We focused on five currently popular second-round regimens and 16 different treatment objectives. Two clinical AIDS experts on our team independently rated the five regimens as first-choice to fifth-choice regimen for each objective and their selections were used as golden standard. We used a genetic algorithm to optimize 20 parameters of our system named AIDS-FDES so that its regimen choices best matched those of the experts individually (i.e., through two different parameters sets). Our preliminary results showed that for the first-choice regimens, the exact agreements between AIDS-FDES and expert A and expert B were 87.5% and 100%, respectively, whereas the mean agreement rate for the five regimens was 77.5% and 80.1%, respectively. For all the five regimens, the agreement within one preference level (i.e., one physician's second choice is another physician's first or third choice), which was an overall agreement measure, for experts A and B was 92.5% and 96.3%, respectively. We also optimized and used just one parameter set to match AIDS-FDES to both the experts simultaneously. The agreement within one preference level for expert A was 90% and 86.3% for expert B. In order to adjust for any agreement likely to occur simply by chance, a weighted Cohen's Kappa was used. The results for the expert's combined selections relative to AIDS-FDES demonstrated that the specialists agreed with the treatment selection made by the computer system with a weighted Cohen's Ka- ppa of 0.78 (95% confidence interval is [0.69, 0.87]), which indicates that the expert's combined agreement with the system's choices (beyond that expected by chance) was importantly improved over that of either expert's agreement with each other