SpeechCARE: dynamic multimodal modeling for cognitive screening in diverse linguistic and speech task contexts - Scorecard - MDSpire

SpeechCARE: dynamic multimodal modeling for cognitive screening in diverse linguistic and speech task contexts

  • By

  • Hossein Azadmaleki

  • Yasaman Haghbin

  • Sina Rashidi

  • Mohammad Javad Momeni Nezhad

  • Ali Zolnour

  • Maryam Zolnoori

  • November 17, 2025

  • 0 min

Share

Clinical Scorecard: SpeechCARE: An Innovative Multimodal Approach for Cognitive Assessment Across Varied Speech and Language Contexts

At a Glance

CategoryDetail
ConditionCognitive impairment including Alzheimer’s Disease and Related Dementias (ADRD) and Mild Cognitive Impairment (MCI)
Key MechanismsMultimodal transformer pipeline integrating acoustic and linguistic speech features with demographic data to detect cognitive impairment from brief speech recordings
Target PopulationAdults over 60 years, including English, Spanish, and Mandarin speakers
Care SettingEarly cognitive impairment screening in clinical and research settings

Key Highlights

  • SpeechCARE achieved an average F1-score of 72.11% and micro AUC of 86.83% on a multilingual test set of 412 participants.
  • The model fuses acoustic (mHuBERT), linguistic (mGTE), and demographic (age) features using an Adaptive Gating Fusion mechanism.
  • SpeechCARE demonstrated strong multilingual generalizability and complements blood-based biomarkers by capturing functional speech deficits.

Guideline-Based Recommendations

Diagnosis

  • Utilize brief speech recordings analyzed by multimodal transformer models to support early detection of cognitive impairment.
  • Incorporate demographic factors such as age to improve predictive accuracy.

Management

  • Employ SpeechCARE as a scalable screening tool to identify individuals requiring further diagnostic evaluation.
  • Use speech-based assessments alongside traditional biomarkers to enhance early diagnosis.

Monitoring & Follow-up

  • Monitor changes in speech patterns over time to track cognitive decline progression.
  • Apply threshold optimization to improve recall for Mild Cognitive Impairment detection.

Risks

  • Be aware of moderate disparities in model performance across language groups, particularly for Spanish speakers.
  • Consider potential biases and validate model performance in diverse populations before clinical implementation.

Patient & Prescribing Data

Older adults with suspected cognitive decline across multiple languages (English, Spanish, Mandarin)

SpeechCARE supports early, non-invasive screening to guide timely intervention and management decisions.

Clinical Best Practices

  • Integrate multimodal speech and demographic data for comprehensive cognitive impairment assessment.
  • Use advanced preprocessing including audio anomaly detection, noise reduction, and speech-task identification to improve data quality.
  • Leverage transformer-based models fine-tuned on aggregated multilingual datasets to enhance generalizability.
  • Complement speech-based screening with established biomarkers and clinical evaluation for accurate diagnosis.

References

Original Source(s)

Related Content