This study addresses the challenges of confounding effects and interpretability in artificial-intelligence-based medical image analysis. Whereas existing literature often resolves confounding by removing confounder-related information from latent representations, this strategy risks affecting image reconstruction quality in generative models, thus limiting their applicability in feature visualization. To tackle this, we propose a different strategy that retains confounder-related information in latent representations while finding an alternative confounder-free representation of the image data. Our approach views the latent space of an autoencoder as a vector space, where imaging-related variables, such as the learning target (t) and confounder (c), have a vector capturing their variability. The confounding problem is addressed by searching a confounder-free vector which is orthogonal to the confounder-related vector but maximally collinear to the target-related vector. To achieve this, we introduce a novel correlation-based loss that not only performs vector searching in the latent space, but also encourages the encoder to generate latent representations linearly correlated with the variables. Subsequently, we interpret the confounder-free representation by sampling and reconstructing images along the confounder-free vector. The efficacy and flexibility of our proposed method are demonstrated across three applications, accommodating multiple confounders and utilizing diverse image modalities. Results affirm the method’s effectiveness in reducing confounder influences, preventing wrong or misleading associations, and offering a unique visual interpretation for in-depth investigations by clinical and epidemiological researchers. The code is released in the following GitLab repository: https://gitlab.com/radiology/compopbio/ai_based_association_analysis . • Propose a novel deep-learning method for the correction of confounders, which is crucial to avoid misleading findings in medical imaging association analysis. • Introduce geometric insights of confounders into the latent space of an autoencoder, and subsequently propose a correlation-based loss as a novel solution for confounder correction via vector orthogonalization. • Enable semantic feature interpretation in confounder-free AI prediction models, and easily handle multiple categorized or continuous confounders. • Performance evaluation of our pipeline on three applications, accommodating multiple confounders and utilizing diverse image modalities.

BACKGROUND: Establishing collaborations between cohort studies has been fundamental for progress in health research. However, such collaborations are hampered by heterogeneous data representations across cohorts and legal constraints to data sharing. The first arises from a lack of consensus in standards of data collection and representation across cohort studies and is usually tackled by applying data harmonization processes. The second is increasingly important due to raised awareness for privacy protection and stricter regulations, such as the GDPR. Federated learning has emerged as a privacy-preserving alternative to transferring data between institutions through analyzing data in a decentralized manner. METHODS: In this study, we set up a federated learning infrastructure for a consortium of nine Dutch cohorts with appropriate data available to the etiology of dementia, including an extract, transform, and load (ETL) pipeline for data harmonization. Additionally, we assessed the challenges of transforming and standardizing cohort data using the Observational Medical Outcomes Partnership (OMOP) common data model (CDM) and evaluated our tool in one of the cohorts employing federated algorithms. RESULTS: We successfully applied our ETL tool and observed a complete coverage of the cohorts' data by the OMOP CDM. The OMOP CDM facilitated the data representation and standardization, but we identified limitations for cohort-specific data fields and in the scope of the vocabularies available. Specific challenges arise in a multi-cohort federated collaboration due to technical constraints in local environments, data heterogeneity, and lack of direct access to the data. CONCLUSION: In this article, we describe the solutions to these challenges and limitations encountered in our study. Our study shows the potential of federated learning as a privacy-preserving solution for multi-cohort studies that enhance reproducibility and reuse of both data and analyses.

Computer-aided diagnosis of dementia using a support vector machine (SVM) can be improved with feature selection. The relevance of individual features can be quantified from the SVM weights as a significance map (p-map). Although these p-maps previously showed clusters of relevant voxels in dementia-related brain regions, they have not yet been used for feature selection. Therefore, we introduce two novel feature selection methods based on p-maps using a direct approach (filter) and an iterative approach (wrapper). To evaluate these p-map feature selection methods, we compared them with methods based on the SVM weight vector directly, t-statistics, and expert knowledge. We used MRI data from the Alzheimer's disease neuroimaging initiative classifying Alzheimer's disease (AD) patients, mild cognitive impairment (MCI) patients who converted to AD (MCIc), MCI patients who did not convert to AD (MCInc), and cognitively normal controls (CN). Features for each voxel were derived from gray matter morphometry. Feature selection based on the SVM weights gave better results than t-statistics and expert knowledge. The p-map methods performed slightly better than those using the weight vector. The wrapper method scored better than the filter method. Recursive feature elimination based on the p-map improved most for AD-CN: the area under the receiver-operating-characteristic curve (AUC) significantly increased from 90.3% without feature selection to 92.0% when selecting 1.5%-3% of the features. This feature selection method also improved the other classifications: AD-MCI 0.1% improvement in AUC (not significant), MCI-CN 0.7%, and MCIc-MCInc 0.1% (not significant). Although the performance improvement due to feature selection was limited, the methods based on the p-map generally had the best performance, and were therefore better in estimating the relevance of individual features.