Introduction to Age Prediction Models and Mental Health Disorders
by Lilo Ryan
We’ve all heard about how machine learning models are biased in many different aspects— from race to gender, but does that hold for mental health disorders? Using publicly available tools, I want to test how the model predicts age based on a comparison with two datasets of subjects performing the same tasks: a control group and a group of people with schizophrenia.
Considering the stigma surrounding mental health disorders, it is valuable to examine how the model predictions are altered based on an existing disorder. Schizophrenia causes studied impacts to the brain and does impact other functioning modalities. It is a disorder that causes hallucinations and psychosis. The neurological basis of the disorder is the dysfunction of dopamine in the brain, though there are other binding proteins that are seen to be functioning differently in the brains of people with schizophrenia. Commonly, people with mental health disorders are not treated the same as neurotypical people in terms of medical treatment and in society— so does that hold true for a model as well?
Age prediction is a valuable tool because it can help identify early signs of cognitive decline. An older prediction for the brain age indicates that stronger methods are needed to be employed to try to prevent more detrimental impacts. In order to identify age, the fMRI data used can be examined for activation in different regions. In past studies, a decrease in activity in the inferior temporal cortex has been found to be indicative of increased age (Luvsannyam et. al). There is a full dataset of schizophrenic individuals with a control set of neurotypical individuals performing shape completion tasks. The data is open source, so it provides a valuable opportunity to investigate with a model. The number of subjects is a bit lower than desired for a model trained from scratch, but I will fortunately be using a pretrained model.
For the models themselves, they range from CNNs to regression models. The model that I am likely using is a SFCN, or a Simple Fully Connected Network. The different publicly available models are indicative that there is a lot of innovation and testing going on in the field.
In terms of existing literature, there are many models tested on datasets with Alzheimer’s patients and correct prediction of their ages. Most models are only trained on people with no cognitive impairment, but they are still able to predict the age of the subjects (Lee, et.al 2022). They successfully also predicted the age of Alzheimer’s patients, which is indicative that there is a neurological trend that the model was able to pick up on.
There have been several studies comparing the different open-source, pretrained models. The most accurate, going off the diagram above, is pyment, which is most likely to be the model used for my investigation. Doerfel, et.al were testing datasets with people with no cognitive impairments. The diagram shown above indicates the accuracy of different open-source models. The boxplot on the top indicates the error in the model in terms of the predicted minus real age.
Pyment is a model that is open source and pretrained. It was trained on heterogeneous tasks and should be able to intake all fMRI data. I will be using their simple fully connected network that works based on regression methods. This will be valuable to use because it has been shown in multiple other papers how successful it is in predicting ages of people with Alzheimer’s and without.
I will run statistical tests on the distances between actual ages and predictions that the model makes. I suspect that the model might perform worse in estimating the age of people with schizophrenia because it will not have seen any training data with neurodiverse individuals— it might be that the model estimates that people with schizophrenia are older than neurotypical individuals. This would be consistent with a study that found that the amount of gray matter in the brain in people with schizophrenia indicated a brain age of over 5 years older than in neurotypical individuals (Shahab et. al).
Hopefully, this investigation will make it more clear how valuable diverse training datasets are.
References
- Dörfel, Ruben P., et al. “Prediction of brain age using structural magnetic resonance imaging: A comparison of accuracy and test–retest reliability of publicly available software packages.” Human Brain Mapping, vol. 44, no. 17, 16 Oct. 2023, pp. 6139–6148, https://doi.org/10.1002/hbm.26502.
- Lee, Jeyeon, et al. “Deep learning-based brain age prediction in normal aging and dementia.” Nature Aging, vol. 2, no. 5, 9 May 2022, pp. 412–424, https://doi.org/10.1038/s43587-022-00219-7.
- Luvsannyam, Enkhmaa, et al. “Neurobiology of schizophrenia: A comprehensive review.” Cureus, 8 Apr. 2022, https://doi.org/10.7759/cureus.23959.
- Shahab, Saba, et al. “Brain structure, cognition, and brain age in schizophrenia, bipolar disorder, and healthy controls.” Neuropsychopharmacology, vol. 44, no. 5, 20 Dec. 2018, pp. 898–906, https://doi.org/10.1038/s41386-018-0298-z.