Researchers at MIT (Massachusetts Institute of Technology) have built a model that can digest multiple types of a patient’s health data to assist doctors to make decisions with partial information. The process entails prognosticating variables of interest like disease risk, from known variables, including symptoms, biometric data, lab tests, and body scans.
Though, patient data can come from many various sources and is often incomplete. For instance, partial information from health surveys about physical and mental well-being, saturated with highly complex data comprising measurements of heart or brain function. So, utilizing Machine Learning to interpret all available data could assist doctors to better diagnose and treat patients. But most models can’t control the highly composite data. MIT researchers, in a paper being presented at the AAAI Conference on Artificial Intelligence next week, define a single neural network that takes as input both simple and highly complex data. Then the network, using the known variables, can fill in all the missing variables. In a patient’s electrocardiography (ECG) signal that measures heart function and self-reported fatigue level, the model can envisage a patient’s pain level and report correctly that patients might not remember.
The network functions thru basting together several submodels, each tailored to specify a specific relationship among variables. These submodels share data as they make forecasts, and ultimately output a prognosticated target variable. The researchers programmed their network to utilize both the traditional method and backpropagation throughout testing. Backpropagation is basically taking a variable output, then predicting an input from that output, and transmitting the input value backward to a preceding node. This formulates a network where all submodels are working together and co-dependent on one another, to output a target probability.