Explainable brain age predictions to support precision diagnosis in neuropsychiatry

1938 

Abstract Submission 

Esten Leonardsen¹, Thomas Wolfers², Lars Westlye³, Yunpeng Wang¹

¹University of Oslo, Oslo, Oslo, ²University of Tübingen, Tübingen, Tübingen, ³Norwegian Centre for Mental Disorders Research (NORMENT), Oslo University Hospital, Oslo, Norway

Esten Leonardsen  
University of Oslo
Oslo, Oslo

Thomas Wolfers  
University of Tübingen
Tübingen, Tübingen

Lars Westlye  
Norwegian Centre for Mental Disorders Research (NORMENT), Oslo University Hospital
Oslo, Norway

Yunpeng Wang  
University of Oslo
Oslo, Oslo

Over the last decade, brain age has emerged as a promising biomarker to encode generalized brain health based on magnetic resonance imaging (MRI) data (Franke & Gaser, 2019). A multitude of studies have shown reliable increases in the brain age gap (BAG), encoding the difference between apparent brain age and chronological age, in patient cohorts across a multitude of neurological and mental disorders (Kaufmann et al., 2019). However, despite its evident utility, clinical translations of brain age models are lacking. One contributing reason for this is the abstract nature of the measure making it hard to understand exactly what it comprises. Furthermore, while being composite makes it sensitive towards detecting deviations across age-related processes, it also potentially lacks the specificity necessary to support clinical decisions for individual patients. Both issues could be alleviated with explainable artificial intelligence. Here, the brain age predictions originating from a machine learning model would be supplemented with an explanation of what caused the prediction in each individual case. One technique for procuring such explanations is Layerwise Relevance Propagation (LRP), commonly used alongside Convolutional Neural Networks (CNNs), to generate heatmaps indicating how regions of an image contribute to the predictions (Bach et al., 2015). In the brain age case, this would allow for precise inference of how different regions in the brain contribute to its overall apparent age. These heatmaps would be richer than the singular BAGs and could facilitate precision diagnosis of neurological and mental disorders.

Here we trained CNNs to predict brain age using 114,289 structural MRIs from 83,401 participants with ages spanning from 3 to 97 years old. All models were variants of the state-of-the-art Simple Fully Convolutional Network architecture, tailored specifically towards procuring informative explanations. On top of the models, we implemented LRP to constitute a fully explainable pipeline for brain age predictions. We employed the pipeline in nine patient cohorts with neurological and mental disorders to investigate whether the heatmaps supported better differentiation of patients and controls than BAGs. Furthermore, we associated the heatmaps with measures of disease severity within the disorders, to evaluate whether they could contribute towards fine-grained stratification of the patients.

Our explainable pipeline was able to predict brain age accurately (MAE=2.92), also in test data coming from sites unseen by the model during training (MAE=4.70). For four out of nine disorders we were able to use BAG to differentiate cases from controls meaningfully (AUC>0.5, p<0.05). For five out of nine disorders, using the heatmaps instead of BAGs significantly improved the predictive performance (p<0.05). Dementia and multiple sclerosis offered the best case-control differentiation both using BAG (AUC=0.75 and 0.63 respectively) and heatmaps (AUC=0.81 and 0.87), but the difference in gain was substantial, potentially indicating differences in the importance of localizing age-related structural anomalies in the two disorders. In both schizophrenia and bipolar disorder using the heatmaps resulted in predictive performance above chance for case-control differentiation (AUC=0.64 and 0.60), but we were unable to use the heatmaps to differentiate between the patient groups (p>0.05), potentially indicating overlap in the structural anomalies in the two disorders.

Brain age has the potential to contribute transformatively towards enabling precision medicine in neurological and mental disorders. Our explainable brain age pipeline contributes towards that goal via establishing trust with potential users of the technology by providing explanations of model behaviour, and by increasing its sensitivity through precise localization of age-related structural anomalies.

Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s)

Neurodevelopmental/ Early Life (eg. ADHD, autism)

Psychiatric (eg. Depression, Anxiety, Schizophrenia)

Classification and Predictive Modeling ²

Methods Development ¹

Aging

Machine Learning

STRUCTURAL MRI