Speech-Based Artificial Intelligence for Detection of Dementia in Danish Patients (DetectAI)

Background Dementia is a growing public health challenge, and early and accurate diagnosis is essential for effective care and potential future disease-modifying treatments. Current diagnostic pathways are resource-intensive and associated with long waiting times. Speech reflects cognitive functioning, and recent international studies have shown that machine learning models can detect dementia-related patterns in speech recordings with promising accuracy. This study aims to develop a speech-based deep learning model in a Danish setting, providing a non-invasive and scalable screening tool for use in primary care.

Study Design and Sampling Methods

This is an observational, cross-sectional study. Participants are recruited using two different sampling strategies corresponding to two artificial intelligence (AI) model development tracks:

Track A (Model A) - Retrospective case-control sampling:

This track addresses a focused diagnostic task: identification of Mild Cognitive Impairment (MCI). Participants are patients with a recent diagnosis from the memory clinic at Region Zealand University Hospital (ZUH). Sampling uses convenience sampling prioritizing patients who live close to the hospital, as data collection occurs during home visits. Patients with more recent diagnoses are prioritized to minimize the risk that participants have progressed to a new disease stage since diagnosis (e.g., from MCI to dementia).

Track B (Model B) - Prospective consecutive sampling:

This track uses prospective inclusion of newly referred patients to the memory clinic without pre-selection by diagnosis, reflecting a real-world clinical screening population. All eligible, consenting patients are included consecutively at their first clinic visit, before final diagnosis is established.

Model Development Following Best Practice Guidelines The study follows TRIPOD-AI (Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis - Artificial Intelligence) and PROBAST-AI (Prediction model Risk Of Bias ASsessment Tool - Artificial Intelligence) guidelines for developing and validating clinical prediction models.

Key methodological features include:

Transparent model development: All preprocessing steps, feature extraction methods, model architectures, and hyperparameters will be documented Robust validation strategy: Data will be split into training, validation, and hold-out test sets for in-depth internal validation.

Minimizing bias: Participant selection, predictor measurement, outcome determination, and statistical analysis are designed to minimize bias according to PROBAST-AI domains Clinically relevant performance metrics: Sensitivity, specificity, area under the receiver operating characteristic curve (AUC-ROC), positive and negative predictive values, and calibration Interpretability: Feature importance analysis to understand which speech characteristics contribute to predictions

Data Collection Speech data is collected through structured tasks including picture description, verbal fluency tests, story recall, and spontaneous speech. Audio is recorded using standardized equipment with quality control checks. Clinical diagnoses are established by experienced clinicians at the memory clinic following international diagnostic criteria.

Inclusion criteria

Model A (patient participants)

• Age > 50 years
• Fluent in Danish
• Minimum of 7 years of schooling
• A diagnosis of either MCI or AD, given at the SUH memory clinic within 6 months before enrollment

Model A (cognitively healthy controls)

• Age > 50 years
• Fluent in Danish
• Minimum of 7 years of schooling

Model B:

• Age > 50 years
• Fluent in Danish
• Minimum of 7 years of schooling

Exclusion criteria

Model A:

Patients:

• Significantly impaired vision or hearing (to the extent that the patient cannot participate in the AI analysis)
• MMSE score < 16
• Concomitant diagnoses which are expected to influence cognitive impairment (eg. depression)
• Patients unable to give consent
• Patients with alcohol consumption >21 standard alcohol units per week
• Any history of speech or language impairment predating the current condition

Cognitively healthy controls:

• Significantly impaired vision or hearing (to the extent that the patient cannot participate in the AI analysis)
• MMSE < 26 and ACE < 90
• Clinical, laboratory, or neuroradiological findings that could affect cognitive functions
• Known diseases which are expected to impair cognitive functions
• Any history of speech or language impairment predating the current condition
• Patients with alcohol consumption >21 standard alcohol units per week.

Model B:

• Significantly impaired vision or hearing (to the extent that the patient cannot participate in the AI analysis)
• MMSE score < 16
• Patients unable to give consent
• Patients with concomitant psychosis or severe psychiatric comorbidities other than depression
• Any history of speech or language impairment predating the current condition