Decision tree-based classification as a support to diagnosis in the Alzheimer's disease continuum using cerebrospinal fluid biomarkers: insights from automated analysis

Abstract

Objective: Cerebrospinal fluid (CSF) biomarkers add accuracy to the diagnostic workup of cognitive impairment by illustrating Alzheimer’s disease (AD) pathology. However, there are no universally accepted cutoff values for the interpretation of AD biomarkers. The aim of this study is to determine the viability of a decision-tree method to analyse CSF biomarkers of AD as a support for clinical diagnosis. Methods: A decision-tree method (automated classification analysis) was applied to concentrations of AD biomarkers in CSF as a support for clinical diagnosis in older adults with or without cognitive impairment in a Brazilian cohort. In brief, 272 older adults (68 with AD, 122 with mild cognitive impairment [MCI], and 82 healthy controls) were assessed for CSF concentrations of Aβ 1-42 , total-tau, and phosphorylated-tau using multiplexed Luminex assays; biomarker values were used to generate decision-tree algorithms (classification and regression tree) in the R statistical software environment. Results: The best decision tree model had an accuracy of 74.65% to differentiate the three groups. Cluster analysis supported the combination of CSF biomarkers to differentiate AD and MCI vs. controls, suggesting the best cutoff values for each clinical condition. Conclusion: Automated analyses of AD biomarkers provide valuable information to support the clinical diagnosis of MCI and AD in research settings.

Machine learning in EP research: New tools for old problems

“…Decision tree algorithms may be favored for openness, using decision-making branches assessing the cut-off values of key predictor variables. 8 Often inspection of which combinations of features that ML algorithms select can provide useful insights, potentially highlighting previously under-appreciated features and patterns in the data.…”

“…selected a decision tree algorithm (with a practically chosen limit of three decision branches), which performed favorably in accuracy, as well as providing full transparency of the decision-making process. 8,10 Selected from 128 input variables, the four-most important input-features used by this algorithm were: aVF/II R-wave ratio, V2S/V3R index, QRS-amplitude in V3, and the R-Wave deflection slope in lead V3. Using these metrics, the performance of this algorithm was extremely high, with an accuracy of 94.4%, precision of 91.5%, recall of 100%, and an F1-score of 0.96.…”

“…Apart from the goal of maximizing accuracy, the choice of classifier algorithm may also depend on transparency requirements, that is, if it is acceptable for the algorithm to function as a black box or if easy interpretation of the reasoning behind decisions is required. Decision tree algorithms may be favored for openness, using decision‐making branches assessing the cut‐off values of key predictor variables 8 . Often inspection of which combinations of features that ML algorithms select can provide useful insights, potentially highlighting previously under‐appreciated features and patterns in the data.…”

“…Upon comparing several classifier algorithms, Shimojo et al selected a decision tree algorithm (with a practically chosen limit of three decision branches), which performed favorably in accuracy, as well as providing full transparency of the decision‐making process 8,10 . Selected from 128 input variables, the four‐most important input‐features used by this algorithm were: aVF/II R‐wave ratio, V2S/V3R index, QRS‐amplitude in V3, and the R‐Wave deflection slope in lead V3.…”

Artificial intelligence for dementia prevention

“…These methods can be further split into regression or classification methods, where the dependent variable to be predicted is either numerical (continuous) or categorical 23 . Supervised ML methods such as logistic regression, random forest, SVMs, gradient boosting, deep learning, and decision trees can be utilized for clinical risk prediction 24 —for example, where dementia diagnostic status is known in combination with features such as demographics, imaging, biomarkers, genetics, comorbidities, symptoms, medication use, and other health indicators are used to build models useful for primary and secondary prevention 13,25–27 . Supervised methods have also been developed to classify biomarker data associated with dementia, 28 and these models can also be trained on neuroimaging data such as magnetic resonance imaging (MRI) or positron emission tomography (PET) scans to classify brain images as healthy or indicative of dementia‐related abnormalities 29 .…”