DiAbot: A Conversational AI System Coupling Large Language Models with an Interpretable Decision Tree for CDR-Style Dementia Screening

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Alzheimer’s disease and related dementias are projected to affect more than 150 million people worldwide by 2050. Early staging with validated instruments such as the Clinical Dementia Rating (CDR) scale is essential for timely intervention, yet access to clinician-administered CDR assessment remains constrained by workforce, time, and geographic barriers. This study complements a previously published machine learning pipeline for Alzheimer’s disease prediction by addressing the downstream task of dementia staging. Because the global CDR score is already derived from the six sub-domain ratings through an established rule-based procedure, the contribution reported here lies not in discovering that mapping but in encoding it in a transparent, deployable form: an explainable decision tree classifier embedded in DiAbot, a large-language-model-fronted conversational system that supports self-administered CDR-style assessment. We extracted 13,453 CDR records from the Alzheimer’s Disease Neuroimaging Initiative (ADNI), removed administrative variables, invalid entries, and missing rows (final n = 13,290), and trained decision tree classifiers under two impurity criteria, Information Gain and Gini Index, using a 70/30 stratified record-level hold-out and ten-fold stratified record-level cross-validation. This classifier-level evaluation uses the six domain scores as recorded during ADNI’s clinician-administered assessment, not scores elicited by the DiAbot chatbot; the trained classifier was separately embedded in a web application in which a prompt-engineered large language model conducts a CDR-style interview and normalizes responses to ordinal domain scores, but the end-to-end accuracy of that full conversational pipeline (chatbot elicitation through to final CDGLOBAL) has not yet been measured, and is not what the headline accuracy figures below report. The Information Gain Decision Tree reproduced the established mapping from the six CDR sub-domain scores to the CDGLOBAL with 99.86% accuracy under the record-level hold-out protocol (matching macro-averaged precision, recall, and F1-score), with a ten-fold record-level cross-validated mean of 99.81% (SD 0.07); this result represents fidelity to the established CDR scoring rule rather than independent dementia-diagnosis accuracy. Gini-based trees performed almost identically (99.79% hold-out, 99.74% cross-validated). Memory dominated feature importance, consistent with its role as the primary domain in the official CDR scoring algorithm. Residual misclassifications were confined to adjacent CDR stages. Because the CDGLOBAL is deterministically derived from the six sub-domain scores, these figures should be read throughout as evidence of high-fidelity reconstruction of the established CDR scoring relationship, not as general dementia-diagnosis accuracy comparable to imaging- or biomarker-based classifiers; further, participant-independent generalization remains unverified under the record-level protocol evaluated here. An interpretable classifier embedded in a conversational front-end can nonetheless make standardized CDR-style staging more widely accessible while preserving clinical inspectability; the resulting system is positioned as a screening-stage adjunct to, and not a replacement for, clinician-administered CDR assessment.