TY  - CONF
AU  - Komeyer, Vera
AU  - Patil, Kaustubh
AU  - Reuter, Martin
AU  - Wolfers, Thomas
AU  - Li, Jingwei
TI  - Data leakage in machine learning: A conceptual take
M1  - FZJ-2024-07535
PY  - 2024
AB  - Symposium:Machine learning (ML) and artificial intelligence (AI) are increasingly being applied to study how individual differences in the brain can manifest as distinct psychiatric illnesses. These models can help us establish the neural correlates of mental distress and predict individual-level diagnosis, symptoms, trajectories, and treatment responses. To realize the full potential of these models it is important to recognize their data requirements as well as biases in data and modeling choices that can limit applicability and insights provided by the models. Biases in these models and data can lead to inaccurate and unfair predictions and overlook individual variations, which can have serious consequences for patients and they can perpetuate and amplify existing health disparities and inequalities. These biases may arise from methodological choices including the neuroimaging modality and state, behavioral phenotypes, data transformation, sample size, population, and modeling pipelines. It is crucial to carefully evaluate the risks associated with AI/ML-based modeling such as biases and develop strategies to identify and mitigate them. In doing so we can improve the accuracy, fairness, and reliability of the predictions and ensure that they benefit all patients equally. This symposium will discuss opportunities and challenges related to application of AI/ML in neuroimaging data from both applied and conceptual perspectives.Data Leakage Talk:ML's popularity stems from the promise of sample-level prediction using high dimensional data. However, if not properly implemented and evaluated, data-leakage in ML pipelines may result in overoptimistic performance estimates and fail to generalize to new data. In this talk I will discuss data-leakage associated challenges and remedies.
T2  - DGKN
CY  - 6 Mar 2024 - 9 Mar 2024, Frankfurt am Main (Germany)
Y2  - 6 Mar 2024 - 9 Mar 2024
M2  - Frankfurt am Main, Germany
LB  - PUB:(DE-HGF)6
UR  - https://juser.fz-juelich.de/record/1034781
ER  -