I have chosen for my disertation thesis to diagnose people having AD. I have used the dataset provided by the dementiabank.

My question it is, do you think it is too simple? I have achieved results close to state of the art but did not beat them. 89% vs 95%claimed by the papers in SoTA.

Do you have any ideeas of what to do in order to be more complex? I have also did some transfer learning which improved the results very much.

  • 2
    $\begingroup$ Well... 89 vs 95 % is not that close, relatively. $\endgroup$ Aug 31 at 8:52
  • $\begingroup$ Be careful in case you're using accuracy: if you have 89% negative cases in the data, achieving 89% accuracy is trivial (majority baseline) and your model is not doing anything interesting. In case the dataset is imbalanced you should probably use F1-score. Another thing to check is whether you have overfitting: if yes, then there's definitely room for improvement. $\endgroup$
    – Erwan
    Aug 31 at 9:04
  • $\begingroup$ The baseline set by the challenge is 75% accuracy. The 89% acc is actually k fold acc achieved on a dataset which contains equal nr of positive and neg samples $\endgroup$
    – Dinu Mihai
    Aug 31 at 9:11
  • $\begingroup$ Well, I think if your intuition is that it is too simple, it probably is. And if you want to pursue a phd, you should look for areas that genuinely have room for improvement, not try to artificially make a simple problem more complex. Anyway, what does your (potential) supervisor think about this? $\endgroup$
    – buddemat
    Sep 2 at 16:53
  • $\begingroup$ My supervisor proposed me this thesis at the begining of year and I have accepted that. At the beginning I said that it is a very interesting thesis since I tought that this project can have a great impact in the individuals life. After I finished my code and looked back at what I did I have this impression that everything I did was not that hard. I am thinking however at developing an app which could diagnose Alzheimer and suggest individuals to go and check themselves eventually maybe. But now I don't have time to do that anymore. $\endgroup$
    – Dinu Mihai
    Sep 5 at 19:32


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.