Improved Motor Neurone Disease Prognosis Prediction with Multimodal Data Fusion
Poster No:
1386
Submission Type:
Abstract Submission
Authors:
Florence Townend1, James Chapman1, Ayodeji Ijishakin1, Federica Agosta2,3, Edoardo Spinelli2,3, Silvia Basaia2, Yuri Falzone3, Paride Schito3, Massimo Filippi2, Andrea Malaspina4, James Cole1
Institutions:
1Centre for Medical Image Computing, University College London, London, UK, 2Neuroimaging Research Unit, Division of Neuroscience, IRCCS San Raffaele Scientific Institute, Milan, Italy, 3Neurology Unit, IRCCS San Raffaele Scientific Institute, Milan, Italy, 4UCL Queen Square Institute of Neurology, University College London, London, UK
First Author:
Co-Author(s):
Federica Agosta
Neuroimaging Research Unit, Division of Neuroscience, IRCCS San Raffaele Scientific Institute|Neurology Unit, IRCCS San Raffaele Scientific Institute
Milan, Italy|Milan, Italy
Neuroimaging Research Unit, Division of Neuroscience, IRCCS San Raffaele Scientific Institute|Neurology Unit, IRCCS San Raffaele Scientific Institute
Milan, Italy|Milan, Italy
Edoardo Spinelli
Neuroimaging Research Unit, Division of Neuroscience, IRCCS San Raffaele Scientific Institute|Neurology Unit, IRCCS San Raffaele Scientific Institute
Milan, Italy|Milan, Italy
Neuroimaging Research Unit, Division of Neuroscience, IRCCS San Raffaele Scientific Institute|Neurology Unit, IRCCS San Raffaele Scientific Institute
Milan, Italy|Milan, Italy
Silvia Basaia
Neuroimaging Research Unit, Division of Neuroscience, IRCCS San Raffaele Scientific Institute
Milan, Italy
Neuroimaging Research Unit, Division of Neuroscience, IRCCS San Raffaele Scientific Institute
Milan, Italy
Massimo Filippi
Neuroimaging Research Unit, Division of Neuroscience, IRCCS San Raffaele Scientific Institute
Milan, Italy
Neuroimaging Research Unit, Division of Neuroscience, IRCCS San Raffaele Scientific Institute
Milan, Italy
Introduction:
Motor Neurone Disease (MND) is a multifactorial and heterogeneous neurodegenerative disease with an expected survival time of 3 to 5 years from symptom onset. During diagnosis, a multimodal set of data can be collected, including brain MRI, blood tests, and functional ratings. Accurate prognosis prediction for personalised clinical management is challenging but vital for empowering patients' and families' future planning and for clinical trial design. However, most current MND prognosis tools only use clinical measures, and the potential contribution of brain MRI to prognosis prediction has been under-explored.
We aimed to explore the advantages of using multimodal data for prognosis prediction compared to using unimodal data, by modelling survival time in MND patients using baseline volumes extracted from structural MRI and clinical measures. Further, we evaluated different multimodal data fusion approaches, via the Fusilli Python package (Townend, 2023).
We aimed to explore the advantages of using multimodal data for prognosis prediction compared to using unimodal data, by modelling survival time in MND patients using baseline volumes extracted from structural MRI and clinical measures. Further, we evaluated different multimodal data fusion approaches, via the Fusilli Python package (Townend, 2023).
Methods:
The study utilised two datasets, University College London Queen's Square Institute of Neurology's ALS Biomarkers Study (UK MND CSG) and Ospedale San Raffaele in Milan, Italy, containing clinical information and brain MRI. We included patients (N=110) in our analysis who had passed away and who had T1w or T2w brain MRI scans within 12 months of diagnosis. Patients were categorised into slow- and fast-progressors based on median survival time from diagnosis to death (see Table 1 for patient demographics).
MRI segmentation was performed using SynthSeg, a modality-agnostic deep-learning segmentation tool (Billot, 2021), that enabled mismatched MRI modalities among patients to be included. The resulting 33 extracted volumes were z-score normalised. The clinical features were age, sex, diagnostic delay (the time between symptom onset and diagnosis), and ALSFRS-R, a functional disability rating scale.
We compared 10 multimodal data fusion methods and 2 unimodal methods on their performance in classifying newly diagnosed MND patients as slow- or fast-progressing. Each method was trained with 10-fold cross validation and the training was repeated until the mean AUC (Area Under the receiver-operating characteristic Curve) stabilised over all repetitions. The AUC for each repetition was computed by aggregating validation folds from the cross-validation training and calculating the AUC of the aggregated folds.
Data fusion methods were compared on their mean AUC over the stability repetitions.
MRI segmentation was performed using SynthSeg, a modality-agnostic deep-learning segmentation tool (Billot, 2021), that enabled mismatched MRI modalities among patients to be included. The resulting 33 extracted volumes were z-score normalised. The clinical features were age, sex, diagnostic delay (the time between symptom onset and diagnosis), and ALSFRS-R, a functional disability rating scale.
We compared 10 multimodal data fusion methods and 2 unimodal methods on their performance in classifying newly diagnosed MND patients as slow- or fast-progressing. Each method was trained with 10-fold cross validation and the training was repeated until the mean AUC (Area Under the receiver-operating characteristic Curve) stabilised over all repetitions. The AUC for each repetition was computed by aggregating validation folds from the cross-validation training and calculating the AUC of the aggregated folds.
Data fusion methods were compared on their mean AUC over the stability repetitions.
Results:
Figure 1 shows the performance distributions over the stability repetitions of each fusion model, arranged by mean AUC from highest to lowest. Of the ten fusion models evaluated, four outperformed the extracted volumes unimodal model (AUC=0.74), and eight outperformed the clinical features unimodal model (AUC=0.65). The highest-performing model (AUC=0.80) was the early concatenation fusion model, where extracted volumes and clinical features were concatenated before being input into a fully connected neural network.
The highest AUC of 0.80 shows that there is promise in using multimodal techniques for MND prognosis prediction over standard unimodal approaches. However, to ensure robust performance on external datasets without using stability repetitions, a larger sample size is needed.
The highest AUC of 0.80 shows that there is promise in using multimodal techniques for MND prognosis prediction over standard unimodal approaches. However, to ensure robust performance on external datasets without using stability repetitions, a larger sample size is needed.
Conclusions:
Data fusion of both brain MRI and clinical features led to improved performance, showing the added value of neuroimaging features predicting MND prognosis. Crucially, the choice of data fusion method influenced predictive performance, with some performing better and some performing worse than unimodal methods. This variability in performance highlights the importance of assessing different data fusion methods and the power of the Fusilli toolbox to achieve this. Future work will explore the fusion of clinical features with voxel-wise MRI data and test generalisability in external datasets.
Disorders of the Nervous System:
Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) 2
Modeling and Analysis Methods:
Classification and Predictive Modeling 1
Methods Development
Multivariate Approaches
Novel Imaging Acquisition Methods:
Multi-Modal Imaging
Keywords:
Computing
Data analysis
Degenerative Disease
Design and Analysis
Machine Learning
Modeling
MRI
Multivariate
Segmentation
STRUCTURAL MRI
1|2Indicates the priority used for review
Provide references using author date format
UK MND CSG, “UK MND CSG — ALS Biomarkers Study.” https://www.mndcsg.org.uk/mnd-clinical-research/research-studies/als-biomarkers-study/
Billot, B. (2021). SynthSeg: Domain Randomisation for Segmentation of Brain Scans of any Contrast and Resolution(arXiv:2107.09559). arXiv