Background:  HAART has dramatically reduced the incidence of AIDS-defining illnesses and mortality in HIV-infected individuals, yet these events continue to occur after initiating therapy. Developing and evaluating prediction models for the development of an individual’s first AIDS-defining illnesses or death after HAART remains an important goal. Utilizing extensive data from persons in the Johns Hopkins HIV Clinical Cohort (Baltimore, Maryland), we developed a new prediction model and compared it to published models.

Methods:  Utilizing data from persons who initiated HAART from 1996 through 2004, we constructed proportional hazards prediction models for the development of first AIDS-defining illnesses or death. We assessed discrimination using time-varying area under the receiver operating characteristic curve (AUC). Analyses were conducted on both a composite (death or AIDS-defining illness) and cause-specific competing-risks endpoint. Variables evaluated at the time of HAART initiation (time-stationary) included CD4 count, HIV RNA, total lymphocyte count, hemoglobin, albumin, and creatinine levels, prior AIDS-defining illness, Pneumocystis carinii pneumonia prophylaxis, sex, race, age, injection drug use (IDU), history of heavy alcohol use, heroin or cocaine use, and a medical history of personality disorder, anxiety, depression, schizophrenia, or suicide attempt.

Results:  Of 2961 individuals who initiated HAART 1235 contributed events (772 AIDS-defining illnesses, 463 deaths) over 10,728 person-years of follow-up. Our prediction model gave a mean AUC over 5 years of 0.73. Variables associated with shorter time to first AIDS-defining illness or death included: CD4, HIV RNA, total lymphocyte count, hemoglobin, and albumin levels, age, prior AIDS-defining illness, IDU, PCP prophylaxis, cocaine use, anxiety and depression. This model provided better discrimination at 6 months (AUC = 0.73) than a published prediction model (AUC = 0.68) that used CD4, HIV RNA, anemia, body mass index, age, ARV use prior to HAART, IDU, and prior AIDS-defining illness. Splitting the composite outcome into competing events resulted in similar 6-month discrimination estimates with the exception of candidiasis (AUC = 0.65).

Conclusions:  We have developed a prediction model for AIDS-defining illnesses and death that has improved ability to discriminate these serious events relative to a published algorithm. As our model retained non-laboratory measures, this suggests their importance in predicting HAART response. This framework highlights an important goal for future clinical research of identifying and validating new measures that improve clinical prediction of competing events.