Neurofind: using deep learning to make individualised inferences in brain-based disorders

Abstract

Within precision psychiatry, there is a growing interest in normative models given their ability to parse heterogeneity. While they are intuitive and informative, the technical expertise and resources required to develop normative models may not be accessible to most researchers. Here we present Neurofind, a new freely available tool that bridges this gap by wrapping sound and previously tested methods on data harmonisation and advanced normative models into a web-based platform that requires minimal input from the user. We explain how Neurofind was developed, how to use the Neurofind website in four simple steps ( www.neurofind.ai ), and provide exemplar applications. Neurofind takes as input structural MRI images and outputs two main metrics derived from independent normative models: (1) Outlier Index Score, a deviation score from the normative brain morphology, and (2) Brain Age, the predicted age based on an individual's brain morphometry. The tool was trained on 3362 images of healthy controls aged 20-80 from publicly available datasets. The volume of 101 cortical and subcortical regions was extracted and modelled with an adversarial autoencoder for the Outlier index model and a support vector regression for the Brain age model. To illustrate potential applications, we applied Neurofind to 364 images from three independent datasets of patients diagnosed with Alzheimer's disease and schizophrenia. In Alzheimer's disease, 55.2% of patients had very extreme Outlier Index Scores, mostly driven by larger deviations in temporal-limbic structures and ventricles. Patients were also homogeneous in how they deviated from the norm. Conversely, only 30.1% of schizophrenia patients were extreme outliers, due to deviations in the hippocampus and pallidum, and patients tended to be more heterogeneous than controls. Both groups showed signs of accelerated brain ageing.

Fig. 1

Distribution of age and sex in the train set before and after addressing data imbalance.

Fig. 2

Main steps to use Neurofind.

Fig. 3

Distribution of age and sex in the clinical datasets used for exemplar applications.

Fig. 4. Results from the exemplar applications for AD and SZ.

A Outlier Index Score (OIS) for AD and SZ. B BAG for patients and HC. C, D Median difference between OIS for schizophrenia / Alzheimer’s disease and respective controls for ROIs with a statistically significant difference between the two. Plots show the mean difference between patients and respective controls and the 95% confidence interval (CI) for this difference. The 95% CI was calculated with bootstrapping (1000 repetitions). ^***p ≤ 0.001, ^**p ≤ 0.010, ^*p < 0.05, ns p > 0.05.

Fig. 5. Proportion of patients with overall OIS within the norm, low, medium, or high deviation.

A In the Alzheimer’s disease group, only a small fraction of patients exhibited an overall OIS within the normative range. The percentage of patients classified with deviations increased progressively, with the majority falling into the high deviation category. B In the schizophrenia group, a greater proportion of patients displayed deviations from the norm, with more individuals classified in the medium and high deviation categories than in the low deviation range.

Fig. 6. Spearman’s correlation between regions of interest (ROIs) with significantly larger deviations from the norm in patients and symptom severity.

A In Alzheimer’s disease, greater deviations from the norm in the left hippocampus, left amygdala, and left inferior temporal gyrus were associated with increased disease severity, as indicated by significant negative correlations. B In schizophrenia, larger deviations in the right hippocampus were significantly associated with greater positive and negative symptom severity.

Fig. 7. Within-group mean pairwise cosine similarity for high-level brain regions.

A Alzheimer’s disease patients exhibited greater homogeneity in brain morphology compared to healthy controls, with the highest similarity observed in the ventricles and subcortical regions. B In contrast, Schizophrenia patients were slightly more heterogeneous than their respective controls, with the greatest differences in cosine similarity found in the ventricles and subcortical regions.