This is the step the book cannot sequence for you — the right method depends entirely on the question. What it can do is show you what scholars in this program chose, and why.
Methods in use here¶
Deep learning — the entry point, and the most widely shared method in the program
Graph neural networks — when your data is a network and flattening it into a vector throws away the structure
Foundation models and transformers — pretrain broadly on open data, fine-tune on your task
Transfer learning — the practical answer to a clinical cohort of a few hundred subjects
Normative modelling — per-subject deviation from a healthy reference, instead of a group average
Geometry-aware methods — respecting the curved space connectomes actually live in
Start from someone else’s problem¶
Often faster than starting from the method: find the scholar whose question resembles yours and read what they did.
Browse by research topic — Alzheimer’s, dementia, aging, Parkinson’s, depression, seizures — or by modality. Each hub lists the scholars working on it with a line on their specific angle.
Getting the statistics right¶
Whichever method you pick, the same handful of things decide whether the result holds up:
Prediction vs inference — decide which question you are answering before choosing how to evaluate it
Multiple comparisons — an uncorrected whole-brain map is not a result
Power and sample size — what your cohort can actually detect
Mixed models — for the longitudinal and multi-site designs most cohorts here have
Harmonization and site effects — before pooling anything
Three things that will come up¶
Data leakage — the most common reason a good accuracy figure does not survive contact with new data
Working with imbalanced data — clinical outcome prediction almost always has far more negatives than positives, and accuracy is a misleading metric when it does
Programming obscurities — why the same function gives different answers in MATLAB and Python, and what to do about it. Written because it cost someone here real time
Tools¶
PyTorch Geometric, BrainLM, Brian for spiking network simulation, and the large language model notes.
Next: Share your work.