Deep reinforcement learning for personalized treatment recommendation

One main challenge in precision medicine is to understand common disease at the molecular level to recommend individualized therapies to patients, permitting high efficacy for different and possibly unknown disease subtypes. Another challenge in clinical practice is how to adapt treatment assignments to possible changes in patients' health states and preferences, including previous treatment history.

Formulating a ranking system measuring the efficiency of the drugs as a scoring function that scores each candidate drug at each ranking position. The scores usually represent some relevance, inducing an ordering of the drugs by sorting them in descending order of relevance to form a ranked list at each position. When moving to the next position, according to the dynamic change of the candidate drugs, the context environment, and the corresponding feature vectors, the scores for the remaining drugs may change under the scoring function.