A recent eClinicalMedicine study utilized machine learning (ML) techniques to develop and test a predictive prognostic model (PPM) for early dementia prediction using real-world patient data.
Challenges in diagnosing dementia at an early stage
Researchers predict that the incidence of dementia will increase by three-fold over the next 50 years. Alzheimer’s disease (AD) currently accounts for 60-80% of all dementia cases.
To date, there remains a lack of effective tools for the early diagnosis of dementia. Memory tests are particularly ineffective at the early stage, as they lack sensitivity. Furthermore, most patients cannot access more specific tests such as lumbar punctures for the assessment of cerebrospinal fluid biomarkers, nor positron emission tomography (PET) scans, which are invasive and costly.
Despite recent advancements, artificial intelligence (AI), models developed using ML techniques are also associated with certain limitations. For example, although cohort data is structured, it can lead to generalizability.
About the study
The researchers of the current developed an interpretable and robust PPM that predicts if and how fast patients at early stages of dementia will progress to AD. Early stages of dementia comprise pre-symptomatic or ‘cognitive normal’ (CN) and mild cognitive impairment (MCI).
To demonstrate the clinical utility of the PPM, the researchers trained the system on baseline, non-invasive, and low-cost data. Thereafter, the PPM was tested on real-world out-of-sample patient data and validated against longitudinal diagnoses in real-world data.
Data obtained came from two clinical cohorts as independent test datasets comprising 272 patients, a research cohort from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) with training and validation samples comprising 410 and 609 patients, respectively, as well as the National University of Singapore’s Memory Aging & Cognition Center dataset (MACC) comprising 605 patients.
To predict future cognitive decline at the early stages of dementia using multimodal data, a trajectory modeling approach was adopted based on Generalized Metric Learning Vector Quantization (GMLVQ). The GMLVQ models were trained to distinguish between stable MCI (sMCI) and progressive MCI (pMCI). Patients with sMCI consistently received an MCI diagnosis within a three-year period, whereas those with pCMI progressed to AD within a three-year period.
The training was achieved using Addenbrooke’s Cognitive Examination Revised memory scale (ACE-R memory), Mini-Mental State Examination (MMSE), and grey matter (GM) density from ADNI data.
Study findings
The PPM was associated with a prediction accuracy of 81.7%, specificity of 80.9%, and sensitivity of 82.4% in determining whether individuals with early dementia will remain stable or progress to AD. There was evidence of an interaction between MMSE, GM density, and ACE-R memory, which demonstrates the role of multimodal features in precisely discriminating between sMCI and pMCI.
Training the model with ACE-R memory and MMSE alone delivered similar performance as training with both cognitive and MRI data. The model performed best when multivariate interactions across multimodal data were utilized.
The model-derived prognostic index was clinically relevant for predicting cognitive health trajectories. For two independent datasets, the PPM-derived prognostic index was derived from the baseline data and was significantly different across groups. The index was significantly higher when trained with MRI and cognitive data for multiple test cases such as AD, moderate MCI, mild MCI, or CN3.
Previous studies have reported that up to 35% of dementia cases are misdiagnosed. Importantly, the PPM index demonstrated the potential to reduce the rate of misdiagnoses by training the system on biological data.
The PPM was associated with superior sensitivity and accuracy as compared to typical assessments in clinical practice, logistic regression models, and multivariate regression models. In validation exercises against longitudinal clinical outcomes, PPM robustly predicted whether individuals at early disease stages like MCI would progress to AD or remain stable. The generalizability of the findings across different memory centers is a significant advancement in the field of AI-guided biomarkers for early dementia.
Conclusions
The study findings provide evidence for an interpretable and robust clinical AI-guided approach to detecting and stratifying patients in the early stages of dementia. This marker has a strong potential for adoption in clinical practice due to its validation against multicenter longitudinal patient data across different geographical regions.
Including data from underrepresented groups, incorporating clinical care data to capture comorbidities, and extending the PPM to the prediction of dementia subtypes is needed before this model can be considered a clinical AI tool.
