A Locally-Run, Multimodal Screening Assistant Combining Conversational NLP, Self-Report Questionnaires, and Clock-Drawing Analysis for Low-Cost Early Cognitive-Decline Pre-Screening

Abstract Early detection of cognitive decline remains one of the most persistent challenges in preventive healthcare: validated instruments such as the MMSE or MoCA require a trained clinician, a scheduled appointment, and roughly thirty minutes of structured testing, a combination of requirements that keeps most people from ever being screened until symptoms are already noticeable to others. We present a multimodal, self-administered pre-screening assistant that combines three independent, freely-available signals — (1) conversational linguistic features extracted from spoken responses via offline automatic speech recognition, (2) a structured 12-item self-report questionnaire modeled on the domain structure of the AD8 and Everyday Cognition (ECog) clinical scales, and (3) automated image analysis of a hand-drawn Clock Drawing Test — into a single feature vector scored by a trained classifier. Every component runs locally on consumer hardware using free, open-source libraries and models, with no per-query API cost. We describe the design of each signal, the specific engineering failures discovered and corrected during development and testing, and a Logistic Regression classifier trained and evaluated via leave-one-out cross-validation on a manually generated synthetic dataset of 20 simulated participant records. The final model, built from 19 features spanning all three signals after an explicit check for label leakage, reaches 85% leave-one-out accuracy against an 80% majority-class baseline. We report this result together with its central limitations — the data are manually generated synthetic data rather than recordings from human participants, and training labels were derived from the system's own questionnaire and clock-drawing thresholds rather than independent clinical diagnosis — and argue that the primary contribution of this work is a demonstrated, fully-integrated, explainable multimodal pipeline rather than a validated clinical tool.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-16
DOI
https://doi.org/10.21203/rs.3.rs-11719517/v1
Primary Topic
Dementia and Cognitive Impairment Research
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

A Locally-Run, Multimodal Screening Assistant Combining Conversational NLP, Self-Report Questionnaires, and Clock-Drawing Analysis for Low-Cost Early Cognitive-Decline Pre-Screening

Aryan Kr Samanta, Angelica Dana
4 links
Research Square
Dementia and Cognitive Impairment Research
preprint

A Locally-Run, Multimodal Screening Assistant Combining Conversational NLP, Self-Report Questionnaires, and Clock-Drawing Analysis for Low-Cost Early Cognitive-Decline Pre-Screening

Aryan Kr Samanta, Angelica Dana
preprint en

Abstract

Abstract Early detection of cognitive decline remains one of the most persistent challenges in preventive healthcare: validated instruments such as the MMSE or MoCA require a trained clinician, a scheduled appointment, and roughly thirty minutes of structured testing, a combination of requirements that keeps most people from ever being screened until symptoms are already noticeable to others. We present a multimodal, self-administered pre-screening assistant that combines three independent, freely-available signals — (1) conversational linguistic features extracted from spoken responses via offline automatic speech recognition, (2) a structured 12-item self-report questionnaire modeled on the domain structure of the AD8 and Everyday Cognition (ECog) clinical scales, and (3) automated image analysis of a hand-drawn Clock Drawing Test — into a single feature vector scored by a trained classifier. Every component runs locally on consumer hardware using free, open-source libraries and models, with no per-query API cost. We describe the design of each signal, the specific engineering failures discovered and corrected during development and testing, and a Logistic Regression classifier trained and evaluated via leave-one-out cross-validation on a manually generated synthetic dataset of 20 simulated participant records. The final model, built from 19 features spanning all three signals after an explicit check for label leakage, reaches 85% leave-one-out accuracy against an 80% majority-class baseline. We report this result together with its central limitations — the data are manually generated synthetic data rather than recordings from human participants, and training labels were derived from the system's own questionnaire and clock-drawing thresholds rather than independent clinical diagnosis — and argue that the primary contribution of this work is a demonstrated, fully-integrated, explainable multimodal pipeline rather than a validated clinical tool.

Research Square
Sharda University (IN)
Good health and well-being
Dementia and Cognitive Impairment Research
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon