アイテム詳細

要旨

The relevant HIV data sets used for predicting outcomes of HIV combination therapies suffer from several problems: different treatment backgrounds of the samples, uneven representation with respect to the level of therapy experience and uneven therapy representation. Also, they comprise only viral strain(s) that can be detected in the patients� blood serum. The approach presented in this paper tackles these issues by considering not only the most recent therapies but also the different treatment backgrounds of the samples making up the clinical data sets when predicting the outcomes of HIV therapies. For this purpose, we introduce a similarity measure for sequences of therapies and use it for training separate linear models for predicting therapy outcome for each target sample. Compared to the most commonly used approach that encodes all available treatment information only by specific input features our approach has the advantage of delivering significantly more accurate predictions for therapy-experienced patients and for rare therapies. Additionally, the sample-specific models are more interpretable which is very important in medical applications.