Language-Based Digital Twins for Elderly Cognitive Assistance

Abstract: Digital twins have emerged as a promising paradigm for personalized healthcare, enabling modeling of individual behavior and health trajectories. In cognitive health, early detection of Mild Cognitive Impairment (MCI) remains challenging, where language and conversational patterns serve as non-invasive biomarkers. In this work, we propose a language-based digital twin framework that leverages LLMs to mimic the conversational behavior of elderly individuals by incorporating stylometric cues and contextual metadata. To evaluate fidelity and cognitive consistency, we introduce a multi-head conditional variational autoencoder (cVAE) that jointly measures reconstruction quality and predicts cognitive scores. Experiments on the I-CONECT dataset show that the digital twin preserves identity-specific characteristics and achieves reconstruction and MoCA prediction errors comparable to real data, while outperforming baseline GPT-generated responses. These results highlight the potential of language-based digital twins as a scalable and non-invasive approach for personalized and continuous cognitive health monitoring.

• The paper proposes a novel digital twin architecture that leverages personalized language analysis and stylometric features to detect early cognitive decline.
• It integrates LLM fine-tuning with multimodal data and a cVAE evaluator to accurately replicate participant responses and predict MoCA scores.
• Empirical results show low reconstruction and prediction errors, emphasizing its potential for scalable, non-invasive cognitive monitoring.

Language-Based Digital Twins for Elderly Cognitive Assistance: Technical Analysis

Framework Overview and Motivation

The paper introduces a novel language-based digital twin architecture for personalized cognitive assistance in the elderly. Digital twins, traditionally used in engineering, are adapted for individualized behavioral modeling through real-time integration of multimodal data streams. In the cognitive health domain, early Mild Cognitive Impairment (MCI) detection is critical, but conventional assessments are invasive, infrequent, and expensive. Spoken language patterns—including lexical diversity, pause distribution, and tempo—have recently been shown to possess substantial predictive power as non-invasive biomarkers.

The proposed framework leverages LLMs, specifically GPT-4.1-mini, fine-tuned with participant-specific conversational records and stylometric annotations. These annotations—pause and tempo tokens—serve to encode timing and rhythm, allowing the model to learn temporal speech dynamics closely correlated with cognitive states. Metadata, including participant demographic information, is systematically incorporated for context-aware response generation. The system is visually summarized in (Figure 1).

Figure 1: Overview of the proposed language-based digital twin framework, integrating textual content, stylometric features, and participant metadata for personalized behavioral modeling and cognitive assessment.

Model Architecture and Evaluation Procedure

Digital Twin LLM Construction

Conversational transcripts from the I-CONECT dataset are augmented with stylometric features to capture individual speaking styles, following robust audio reprocessing and speaker diarization protocols for data sanitation. Fine-tuning is performed via supervised learning, with prompts formatted to delineate the mimicry task, user query, and relevant metadata. The LLM is trained to maximize log-likelihood over personalized response generations, aligning output with individual linguistic idiosyncrasies.

Conditional Variational Autoencoder (cVAE) Evaluator

To quantitatively evaluate fidelity and cognitive consistency, the paper employs a multi-head cVAE. The cVAE is conditioned on conversation context and metadata, mapping inputs to a latent space parameterized by $\mu$ and $\log \sigma^2$. The decoder reconstructs participant responses while a parallel head predicts MoCA scores as a proxy for cognitive status. The composite loss function weights reconstruction (MSE), KL divergence, and MoCA prediction, enabling joint assessment of behavioral and cognitive relevance.

Training curves for the digital twin reveal stable convergence and accuracy improvement, as illustrated in (Figure 2).

Figure 2: Normalized loss and accuracy of the digital twin during supervised fine-tuning, indicating robust convergence.

Loss dynamics for the cVAE during training are shown in (Figure 3):

Figure 3: Normalized loss function of the cVAE model for the first 20 epochs; loss consistently decreases, indicating effective learning.

Empirical Results and Analysis

Identity Preservation and Response Fidelity

Classification experiments using SVMs demonstrate high identity detection accuracy for digital twin outputs, closely mirroring real participant responses in both embedding and sentiment spaces. By contrast, standard GPT-generated outputs perform poorly in attribution tasks, underscoring the necessity of fine-tuning and stylometric augmentation.

Cognitive Alignment: MoCA Prediction and Reconstruction Errors

Reconstruction error for participant responses and digital twin outputs remain low ($\sim$0.009), demonstrating the twin’s ability to emulate linguistically and contextually accurate responses. MoCA prediction errors for the digital twin are comparable to those for real data (mostly $<1.1$), whereas raw GPT predictions exhibit a marked deficit ($>3.5$). Thus, the digital twin not only preserves linguistic fidelity but also successfully encodes cognitively relevant information robustly.

Practical and Theoretical Implications

The language-based digital twin paradigm constitutes a significant step forward in both continuous behavioral modeling and scalable cognitive assessment. Practically, the system offers a pathway for real-time, non-invasive MCI screening, potentially mitigating the barriers imposed by traditional diagnostic approaches. Theoretically, this architecture advances digital twin frameworks from population-level physiological modeling toward individualized linguistic and cognitive emulation.

Key implications include:

• Enhanced Clinical Utility: The framework’s ability to jointly capture identity-specific conversational nuances and cognitive signatures enables finely-grained, longitudinal monitoring.
• Scalable Deployment: Integration with telehealth platforms and conversational AI agents can radically expand access to cognitive screening, particularly in socially isolated elderly populations.
• Extension to Multimodal Digital Twins: Incorporation of additional modalities (e.g., audio/video, paralinguistic cues) could enrich behavioral modeling and affective state detection, facilitating early intervention strategies.

Future research should address scalability limitations by expanding evaluation to larger and more heterogeneous cohorts and investigate cross-participant generalization of twin models.

Conclusion

This work presents a rigorous, technically sophisticated framework for language-based digital twins targeting elderly cognitive assistance. Through stylometric augmentation and LLM fine-tuning, the system produces personalized responses with high fidelity and cognitive alignment, as validated by low reconstruction and MoCA prediction errors. The cVAE evaluator facilitates robust joint analysis of linguistic and cognitive metrics. This approach holds promise for scalable, non-invasive continuous cognitive monitoring, and its extension to multimodal data streams is expected to further enhance the depth and utility of individualized digital twins in computational health.