My goal is to predict which depressed patients respond well to a specific treatment. Let's assume I have 2 groups, one has received the treatment ('active group'), the other not ('control group'). It's relatively easy to build a machine learning model which can predict symptom change over time, but how do I identify the factors which predict a positive response to the treatment? Just looking at symptom improvement is not enough, as depressed patients might improve over time without treatment. My guess is that I have to look into interaction features. Any ideas or suggested readings specifically for applying machine learning algorithms to experimental studies would be highly welcome.
Dear Matthias Luethi
Greetings!
From your question, Structural Elucidation and Quantification (SEM Studies) helps to you find out the interactions between the factors.
Secondly, initially you may consider Decision Tree to find out the most influencing factor too.
Dear Matthias Luethi
Generally, Managing machine learning experiments, trials, jobs and metadata using Amazon SageMaker
Kind Regards
Qamar Ul Islam
Dear @Matthias Luethi this link maybe useful
https://link.springer.com/chapter/10.1007%2F978-981-10-7617-6_4
Thank you so much for your answers. My question was less about which algorithm to use or how to train a model, optimize hyperparameters, and evaluate performance. Instead, I am interested in learning more about how to interpret the results. Specifically, I want to understand which patients have profited from treatment (and separate them from patients that have recovered for other reasons). In other words, my question is which personal factors make a patient more likely to respond to treatment. From what I know so far, the key is to look at interactions, between treatment and other predictors.
For example, in regression, interaction terms must be calculated beforehand and entered into the model as predictors. My approach is data-driven, so adding an interaction term for every variable in my dataset doubles the size of an already extensive list of predictors. As a way out, one can use elastic-net regression, because it excludes less relevant predictors from the model (L1 penalization).
But regression might not be the best choice for a lot of predictors anyways, and other algorithms such as Random Forest and XGBoost are likely to perform better. These models are based on decision trees, which can naturally handle interaction of features (no need to include extra interaction predictors). But then the question remains how do I know if treatment made a difference and in which patients? In other words, how do I find out if interactions between treatment and other variables were significant predictors (problem of black-box models)? I don't know too much about this yet, but I am currently looking at model-agnostic interpretation methods, e.g. partial dependence/accumulated local effects plots (Article Visualizing the Effects of Predictor Variables in Black Box ...
). I’d also recommend the free ebook “Interpretable Machine Learning”, which has a section on feature interaction (https://christophm.github.io/interpretable-ml-book/interaction.html).If anybody has suggestions or resources to share, I’d be happy to hear about it.
- Badges
- Science topic