Computational patient-illness and treatment-outcome prediction models are important tools supporting medical professionals. These allow them to support diagnosis, perform effective treatment-matching, and establish preventative care. While, in the past, attaining such models required meticulously crafted expert-knowledge, the modern approach relies on Machine Learning (ML). This approach, in essence, compiles multiple input-output examples (data) from the past, and uses them to create a black-box prediction function. Yet, medical data is often characterized with properties that are a significant challenge for “standard” ML approaches: first, medical data is scarce—we often only have access to “small” or personalized data, while ML typically relies on “big data;” second, this data is often heterogeneous—coming from varied sources in different formats—which can be especially problematic, e.g., for training Neural Networks (NNs).
To overcome these challenges, with this project we aim to establish a novel computational methodology for building prediction models in medical contexts.
The project’s contribution is based on Dr. Elimelech’s cross-disciplinary expertise. Dr. Elimelech’s main field of work is developing efficient planning and learning algorithms for autonomous robots. A recent technique we established, called “learning by abstraction,” allows a robot to improve its planning (i.e., search for actions to execute to achieve its goals) capabilities throughout its lifetime, even given only a few experiences (small data), by making and exploiting generalized conclusions.
With this project, we are interested to see how this technique, or a variation of it, can be used in medical applications, such as learning personalized treatment-outcome prediction models for patients. The suggested technique should be very suitable for such cases, as it can inherently learn from small data and varied-in-resolution and even multi-modal observations. This technique should also be effective for cases in which specifications cannot be formally stated, but examples from an expert can be given.