Prognostic modeling of cognitive decline with confidence quantific | Bruno Hebling Vieira | VIScon23 [Video]

The translational application of precision psychiatry demands uncertainty estimation and calibration. In this talk, we will go through the process of identifying and exploiting opportunities to apply probabilistic machine learning to prediction of future cognitive decline from brain imaging and risk factors, and how this enables confidence quantification and better prognostics.

Prediction of subjective future cognitive decline, usually assessed by the Clinical Dementia Rating (CDR), in elders from appropriate input domains (e.g., brain imaging, genetics, demographics) is usually performed in research applications with regression algorithms that return point estimates. This has achieved great success towards identifying the potential of different variables to better predict clinical outcomes in dementia care. However, translational applications necessarily require uncertainty estimation. In this talk, we will go through the process of identifying certain data patterns that allow us to make better and more meaningful inferences. We will turn a problem that, at first glance, can be well modeled as a regression, into a probabilistic prediction task. Specifically, we identify that predicting across the six categories that compose the CDR sum-of-boxes score with gradient boosted decision trees naturally leads to valid probabilistic predictions. We further investigate where the model succeeds (and fails), and what this might entail for prediction neuropsychiatry and future applications.

Speaker: Bruno Hebling Vieira

https://viscon.ethz.ch
–
https://instagram.com/visconvention
https://linkedin.com/company/visconvention
https://facebook.com/VISconvention