MIT Developing AI To Better Diagnose Cancer

stowie writes: Working with Massachusetts General Hospital, MIT has developed a computational model that aims to automatically suggest cancer diagnoses by learning from thousands of data points from past pathology reports. The core idea is a technique called Subgraph Augmented Non-negative Tensor Factorization (SANTF). In SANTF, data from 800-plus medical cases are organized as a 3D table where the dimensions correspond to the set of patients, the set of frequent subgraphs, and the collection of words appearing in and near each data element mentioned in the reports. This scheme clusters each of these dimensions simultaneously, using the relationships in each dimension to constrain those in the others. Researchers can then link test results to lymphoma subtypes.

Comments

Re:Higher diagnoses

[ColdWetDog]

Probably not - at least in this case. They are looking at a specific form of cancer, lymphoma. Lymphomas do span the gamut from being indolent to extremely aggressive, hence the need for accurate diagnosis, but we have a fairly good idea of what the natural history of each subtype is. This system is not designed to mow through a bunch of clinical data and pop out a 'cancer' diagnosis.

That said, TFA is incredibly poorly written. It is anything but clear WHAT information they are using (pathology slides? DNA samples? Chart notes?) and it is most certainly not AI.

While over diagnosing pre clinical cancers is a concern, this particular methodology won't make that worse. In fact, if it actually does work, it might decrease what are essentially false positive diagnoses by linking the testing component to the natural history of the disease (eg, 'this particular cancer is mostly harmless, don't worry about it much').