Summary: Researchers have developed a new algorithm that integrates Alzheimer’s indicators from MRI measurements and other biomarkers to predict patients at risk of the neurodegenerative disease before symptoms severely affect daily life.
Source: Case Western Reserve.
New machine learning software from Case Western Reserve University shows promise for earlier, more accurate diagnosis of Alzheimer’s disease, according to initial results.
Alzheimer’s disease affects millions of people in the United States and worldwide, and the number of cases is rising as populations age. The condition is an irreversible, progressive brain disorder that gradually destroys memory and cognitive skills. Although no cure exists, several treatments can slow symptom progression for months to years when the disease is detected early. Early diagnosis is therefore critical to help patients maintain independence and to allow clinicians to apply treatments and care strategies sooner.
Researchers at Case Western Reserve have created a machine learning framework designed to improve early detection by combining many different indicators of Alzheimer’s. The program integrates data from structural and functional brain imaging, features derived from the hippocampus, measures of brain glucose metabolism, proteomic and genomic features, and clinical assessments such as mild cognitive impairment (MCI). By evaluating these heterogeneous data sources together, the algorithm aims to provide a more accurate picture of disease risk than any single measurement alone.
“There’s a continuum from healthy aging to mild cognitive impairment to Alzheimer’s,” said Anant Madabhushi, F. Alex Nason Professor II of Biomedical Engineering at Case Western Reserve. “Rather than treating all patients as simply healthy or diseased, we deliberately included mild cognitive impairment, which can be a precursor to Alzheimer’s disease but does not always progress to it.”
In initial testing, the team evaluated the algorithm using Alzheimer’s Disease Neuroimaging Initiative (ADNI) data. The paper and supporting materials report use of patient cohorts drawn from ADNI; in different parts of the report sample sizes of 159 and 149 patients are referenced as the algorithm was evaluated and validated across available datasets.
The researchers call the new approach Cascaded Multi-view Canonical Correlation (CaMCCo). CaMCCo treats each data type as a distinct “view” of the disease and applies a two-stage cascading selection process. In the first stage, the method identifies which features and modalities best separate cognitively normal individuals from those with any impairment. In the second stage, it refines the selection among impaired individuals to distinguish those with mild cognitive impairment from those with full Alzheimer’s disease. The selected views are then fused to build the most informative predictive models for each step.
Madabhushi and colleagues note that their lab has previously shown benefits from integrating diverse data types when identifying and characterizing cancers. CaMCCo applies the same multimodal fusion principle to Alzheimer’s diagnosis, using selective fusion tailored to each classification task rather than combining every available modality indiscriminately.
According to the published results, selective fusion with CaMCCo produced substantially better predictive performance than either single modalities or naive fusion of all modalities together. The authors report that fusing selected modalities yielded a mean area under the receiver operating characteristic curve (AUC) of 0.92, compared with a mean AUC of 0.54 when all modalities were fused without selection. Individual modality AUCs reported in the study varied across modalities, and CaMCCo also achieved stronger positive predictive value for identifying MCI patients (reported PPV: 0.80) compared with other multi-class methods evaluated (PPV values of 0.67 and 0.63 in competing approaches).
These results indicate that selecting and combining the most informative views for each classification step—healthy vs. impaired, then MCI vs. Alzheimer’s—produces a more accurate and clinically useful model. In practice, that means clinicians could receive a clearer, data-driven signal about which patients are at higher risk of progressing toward Alzheimer’s, enabling earlier intervention and more focused monitoring.
The research team plans further validation and refinement using data drawn from multiple sites and larger patient cohorts. Their next steps include running the software in an observational setting alongside clinical workflows so neurologists can see how CaMCCo performs as real patient data are entered. If prospective validation confirms early findings, the investigators anticipate pursuing a clinical trial to evaluate the algorithm’s value in routine care.
Funding: This study was supported by the National Institutes of Health, the National Center for Research Resources, and Department of Defense awards, including Prostate Cancer Synergistic Idea Development Award (PC120857), Lung Cancer Idea Development New Investigator Award, and Prostate Cancer Idea Development Award.
Source: Kevin Mayhood – Case Western Reserve
Original Research: “Cascaded Multi-view Canonical Correlation (CaMCCo) for Early Diagnosis of Alzheimer’s Disease via Fusion of Clinical, Imaging and Omic Features” by Asha Singanamalli, Haibo Wang & Anant Madabhushi. Published in Scientific Reports, August 15, 2017.
Abstract
Cascaded Multi-view Canonical Correlation (CaMCCo) for Early Diagnosis of Alzheimer’s Disease via Fusion of Clinical, Imaging and Omic Features
The addition of mild cognitive impairment (MCI) as a diagnostic category complicates Alzheimer’s Disease (AD) classification because no single biomarker reliably separates all diagnostic groups. Previous fusion studies often failed to consider all diagnostic categories simultaneously and used the same modality set to predict all classes, producing suboptimal fused representations. CaMCCo addresses these issues by providing a cascaded fusion and classification framework that incorporates all diagnostic categories and optimizes classification by selectively combining a subset of modalities at each cascade level. Evaluated on ADNI cohorts for which neurophysiological, neuroimaging, proteomic and genomic data were available, CaMCCo demonstrated that task-specific fusion outperforms both individual modalities and non-selective fusion across multiple metrics, improving mean AUC and MCI prediction performance.