New paper from the Trier team on using Neural Networks to predict remission (or rather RCSI) after psychotherapy. Impressive sample size and interesting results:
Comments
Log in with your Bluesky account to leave a comment
- The best model (TensorFlow) had a PPV of .58, so if the model predicts remission for a patient, they have an 58% change of remitting, compared to a base rate of 43%.
- The NPV was 0.65, so if the model predicts no remission, the chance of not remitting is 65% instead of 57% (the base rate)
- The best model was only minimally better than a generalized linear model. I would suspect that even a linear model including baseline PHQ-0 would have performed not that bad.
There weren't that many predictors in the data set and many of them were short scales of psychopathology.
Finding it odd to refer to the model as "TensorFlow", since it is just a kind of software, but not directly a model. It's like saying "The best model (Stata) ..."
But there are tons of different kinds of neural networks is my point. Transformer architectures, RNNs, CNNs, many different ways. It's like calling random-effects or so a "linear model". Technically true but really weird to say so
As I have no experience with it, I thought using TensorFlow implies some kind of model like a CNN (just like the R function nnet implies a feed-forward network) - well, apparently it doesn’t, thanks for clarifying.
No worries. It is just a general framework with which you can build anything like regular OLS, logistic regression, up to a GPT model or a long-short-term memory for time-series or what have you. But no worries, can happen to everyone. 😊 That's how we learn.
Comments
- The NPV was 0.65, so if the model predicts no remission, the chance of not remitting is 65% instead of 57% (the base rate)
There weren't that many predictors in the data set and many of them were short scales of psychopathology.