Statistical learning methods for bias-aware HIV therapy screening

The human immunode ciency virus (HIV) is the causative agent of the acquired immunode ciency syndrome (AIDS) which claimed nearly 30 million lives and is arguably among the worst plagues in human history. With no cure or vaccine in sight, HIV patients are treated by administration of combinations of antiretroviral drugs. The very large number of such combinations makes the manual search for an e ective therapy practically impossible, especially in advanced stages of the disease. Therapy selection can be supported by statistical methods that predict the outcomes of candidate therapies. However, these methods are based on clinical data sets that are biased in many ways. The main sources of bias are the evolving trends of treating HIV patients, the sparse, uneven therapy representation, the di erent treatment backgrounds of the clinical samples and the di ering abundances of the various therapy-experience levels. In this thesis we focus on the problem of devising bias-aware statistical learning methods for HIV therapy screening predicting the e ectiveness of HIV combination therapies. For this purpose we develop ve novel approaches that when predicting outcomes of HIV therapies address the aforementioned biases in the clinical data sets. Three of the approaches aim for good prediction performance for every drug combination independent of its abundance in the HIV clinical data set. To achieve this, they balance the sparse and uneven therapy representation by using di erent routes of sharing common knowledge among related therapies. The remaining two approaches additionally account for the bias originating from the di ering treatment histories of the samples making up the HIV clinical data sets. For this purpose, both methods predict the response of an HIV combination therapy by taking not only the most recent (target) therapy but also available information from preceding therapies into account. In this way they provide good predictions for advanced patients in mid to late stages of HIV treatment, and for rare drug combinations. All our methods use the time-oriented evaluation scenario, where models are trained on data from the less recent past while their performance is evaluated on data from the more recent past. This is the approach we adopt to account for the evolving treatment trends in the HIV clinical practice and thus o er a realistic model assessment.