Methods: Age Prediction Model for Individuals with Schizophrenia
By Lilo Ryan
This project is made possible by the open source data published in the European Journal of Neuroscience. There are two populations recorded performing the same tasks, which eliminates significant uncertainty in the results. The dataset is largely already processed and is nearly ready to be input to the model.
The shape completion task that was performed on both populations for the data sampling involved seeing if people could determine an illusory or fragmented shape to complete mentally (Figure 1) . The “block arrangement” indicates the way that the stimuli were organized on the screen. In terms of the participants’ physical set-up, they were placed 100 cm away from a scanner with a mirror. The images are referred to as pac-men— the 60 degree chunk out of a circle is indicative of the pac-man shape. The process by which the different signals were acquired can be seen in Figure 1c, where it describes how the participants focused, received instructions, performed the trials for 1.25 minutes, and then focused again.
For the dataset with the bipolar and schizophrenic individuals, the youngest participant is 23 and the oldest is 55. This is well within the range of what the Pyment module is trained on, which means that we don’t have to worry about making any large inferences based on what the model has already received MRI scans for.
The processing is performed as a process called “skull stripping,” which means that the only signals left in the data are the brain signals. All of the non-brain tissue that is captured in the raw data is removed. This is beneficial because it can be assumed that all of the data in the datasets is a real and valid signal from the brain.
The Pyment module is able to be easily tested as a result of being pretrained. The acquisition and discovery of the pretrained model was described in the prior blog post. What is left is mostly just data analysis after testing the dataset on the model. There are many different statistical tests that can be used to determine the model accuracy as well as important summary statistics. The model itself can be easily downloaded from GitHub, where the programmer’s open source code can be found. There were 34,000 samples in the training data, so it is reasonable to assume that the model is reasonably accurate to the general population.
The key summary statistics to observe are the mean and standard deviation of the model’s predicted values minus the actual age of the participants.
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.
- Keane, Brian P, et. al (2024). The Neural Basis of Visual Shape Completion in Schizophrenia and Bipolar Disorder. OpenNeuro. [Dataset] doi: doi:10.18112/openneuro.ds005073.v1.0.0
- Keane, Brian P., et al. “Dorsal attention network activity during perceptual organization is distinct in schizophrenia and predictive of cognitive disorganization.” European Journal of Neuroscience, vol. 57, no. 3, 4 Jan. 2023, pp. 458–478, https://doi.org/10.1111/ejn.15889.