AgeML: a Python package for Age Modelling with Machine Learning made easy.
Poster No:
1170
Submission Type:
Abstract Submission
Authors:
Jorge Garcia Condado1,2, Iñigo Tellaetxe Elorriaga1,2, Asier Erramuzpe1,2,3, Jesus Cortes1,2,3
Institutions:
1Biobizkaia Health Research Institute, Barakaldo, Spain, 2UPV/EHU, Leioa, Spain, 3Ikerbasque, Bilbao, Spain
First Author:
Co-Author(s):
Iñigo Tellaetxe Elorriaga
Biobizkaia Health Research Institute|UPV/EHU
Barakaldo, Spain|Leioa, Spain
Biobizkaia Health Research Institute|UPV/EHU
Barakaldo, Spain|Leioa, Spain
Asier Erramuzpe
Biobizkaia Health Research Institute|UPV/EHU|Ikerbasque
Barakaldo, Spain|Leioa, Spain|Bilbao, Spain
Biobizkaia Health Research Institute|UPV/EHU|Ikerbasque
Barakaldo, Spain|Leioa, Spain|Bilbao, Spain
Jesus Cortes
Biobizkaia Health Research Institute|UPV/EHU|Ikerbasque
Barakaldo, Spain|Leioa, Spain|Bilbao, Spain
Biobizkaia Health Research Institute|UPV/EHU|Ikerbasque
Barakaldo, Spain|Leioa, Spain|Bilbao, Spain
Introduction:
BrainAge models (Franke et al. 2010, Neuroimage) have had success in exploring the relationship between healthy and pathological ageing of the brain. Furthermore, this type of age modelling can be extended to multiple body systems and modelling of the interactions between them (Tian et al 2023, Nature Medicine). However, there is no standard for age modelling. There have been works attempting to describe proper procedures, especially for age-bias correction (de Lange and Cole 2020, Neuroimage: Clinical). In this work we developed an Open-Source software that allows anyone to do age modelling following well-established and tested methodologies for any type of clinical data. Age modelling with machine learning made easy.
Methods:
AgeML is an Open-Source Python library for age modelling which can be found at https://github.com/compneurobilbao/AgeModelling. The package can currently do four types of processing: age modelling and prediction and calculation of age delta (the difference between predicted and chronological age), the association of age delta with lifestyle factors, evaluation of age delta for different clinical groups, and classification of groups based on age deltas.
Age is modelled using classical machine learning algorithms to perform a supervised task of using the provided features to predict subject-wise chronological age using cross-validation. Age bias correction is also applied (de Lange and Cole 2020, Neuroimage: Clinical). The user can specify a covariate, such as gender, to train separate models and provide a file with information about clinical status to train only on controls. After the model is trained on controls it is applied to the other groups. A file is saved with the predicted ages and the age delta for possible further analysis.
The three extra workflows require as input the output from the age modelling workflow. The association of the age deltas with lifestyle factors is done by running a correlation analysis between the deltas and the provided factors. Given a clinical file showing different clinical groups (healthy vs ill, disease subtypes) it computes which groups have deltas that are statistically significantly different. Finally, given two clinical groups and their age deltas, it can train a logistic regressor to classify between each group.
Age is modelled using classical machine learning algorithms to perform a supervised task of using the provided features to predict subject-wise chronological age using cross-validation. Age bias correction is also applied (de Lange and Cole 2020, Neuroimage: Clinical). The user can specify a covariate, such as gender, to train separate models and provide a file with information about clinical status to train only on controls. After the model is trained on controls it is applied to the other groups. A file is saved with the predicted ages and the age delta for possible further analysis.
The three extra workflows require as input the output from the age modelling workflow. The association of the age deltas with lifestyle factors is done by running a correlation analysis between the deltas and the provided factors. Given a clinical file showing different clinical groups (healthy vs ill, disease subtypes) it computes which groups have deltas that are statistically significantly different. Finally, given two clinical groups and their age deltas, it can train a logistic regressor to classify between each group.
·Overview of the 4 workflows of AgeML
Results:
To validate the workflow of this project data were obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database (Petersen et al 2010, Neurology). This included 629 healthy controls (CN), 635 MCI and 208 AD subjects. From the MCI subjects, there were 152 pMCI subjects, those who converted to AD within 3 years from baseline, and 152 sMCI subjects, those who remained MCI for 3 years from baseline, were chosen at random to balance the dataset. This was used to train BrainAge models, show differences in deltas between CN, MCI, AD, pMCI and sMCI, and look at the classification power of the age delta between groups as shown in previous studies (Garcia Condado and Cortes 2023, Alzheimer's & Dementia: DADM).
The software package follows well-established Open Source development practices. It has a continuous integration pipeline to check for PEP8 compliance, unit testing and testing coverage. It is also open source and publicly available to encourage community-driven development, testing and transparency.
The software package follows well-established Open Source development practices. It has a continuous integration pipeline to check for PEP8 compliance, unit testing and testing coverage. It is also open source and publicly available to encourage community-driven development, testing and transparency.
Conclusions:
The objective of AgeML is to standardise procedures, lower the barrier to entry into age modelling and ensure reproducibility. The project is Open-Source to create a welcoming environment and a community to work together to improve and validate existing methodologies. We are actively seeking new developers who want to contribute to growing and expanding the package. Future steps are the implementation and easy access of foundational age models within the package.
Lifespan Development:
Aging 1
Modeling and Analysis Methods:
Classification and Predictive Modeling
Methods Development
Neuroinformatics and Data Sharing:
Workflows
Informatics Other 2
Keywords:
Aging
Computational Neuroscience
Data analysis
Design and Analysis
Machine Learning
Open-Source Code
Open-Source Software
Workflows
1|2Indicates the priority used for review
Provide references using author date format
Franke, K., Ziegler, G., Klöppel, S., & Gaser, C. (2010). Estimating the age of healthy subjects from T1-weighted MRI scans using kernel methods: Exploring the influence of various parameters. NeuroImage, 50(3), 883–892. https://doi.org/10.1016/j.neuroimage.2010.01.005
Garcia Condado, J., Cortes, J. M., & for the Alzheimer’s Disease Neuroimaging Initiative. (2023). NeuropsychBrainAge: A biomarker for conversion from mild cognitive impairment to Alzheimer’s disease. Alzheimer’s & Dementia: Diagnosis, Assessment & Disease Monitoring, 15(4), e12493. https://doi.org/10.1002/dad2.12493
Tian, Y. E., Cropley, V., Maier, A. B., Lautenschlager, N. T., Breakspear, M., & Zalesky, A. (2023). Heterogeneous aging across multiple organ systems and prediction of chronic disease and mortality. Nature Medicine, 29(5), 1221–1231. https://doi.org/10.1038/s41591-023-02296-6