Inferring Affect and Intervention Opportunities for Cancer Survivors from Digital Diaries with Context-Aware LLMs

Paper Title

Inferring Affect and Intervention Opportunities for Cancer Survivors from Digital Diaries with Context-Aware LLMs

Publication Info

• Topic area: Context-aware language models for mental health intervention in cancer survivorship.
• Keywords: Cancer survivors, affective inference, mobile diaries, JITAI, emotion regulation, LLMs, context-aware systems, digital health, sparse text analysis, mental health support.

Background and Problem

• Problem / challenge: Cancer survivors face significant unmet psychosocial needs, with barriers to accessing traditional mental health services. Low-burden methods for monitoring affective states to trigger just-in-time adaptive interventions (JITAIs) are underdeveloped, especially for sparse, introspective mobile diary entries.
• Significance: Addressing mental health challenges in cancer survivorship is critical for long-term well-being and recovery. Enabling timely, context-aware digital interventions can mitigate access barriers, particularly in underserved areas.
• Motivation and related work: Prior work in mobile health and language models has explored affective inference but struggles with sparse text inputs like mobile diaries. Existing methods lack contextual grounding and personalization, limiting their applicability for real-time intervention decisions.

Solution

• Proposed approach: Context-Aware Language Model (CALLM) framework.
• Novelty:
  1. Characterization of contextual cues in ultra-brief mobile diary entries related to affective states, regulation desire, and intervention availability.
  2. Development of CALLM, which integrates similarity-aligned peer cases and short personal trajectories for context-aware inference.
  3. Identification of model-conditioning factors (confidence, diary length, temporal context, short-horizon calibration) that influence inference performance.
• Procedure and key techniques:
  • Diary entries are embedded using text-embedding models and stored in a vector database for similarity-based retrieval.
  • Retrieved peer cases and personal temporal trajectories are incorporated into LLM prompts for affective inference.
  • Predictions are formatted as JSON outputs for downstream use, with balanced accuracy and ROC-AUC metrics guiding evaluation.
  • Experiments include comparisons against baselines (traditional pipelines, fine-tuned transformers, emotion-specialized models, and zero/few-shot LLMs).

Results

• Concrete findings: CALLM achieved balanced accuracy of 72.96% (positive affect), 73.29% (negative affect), 73.72% (regulation desire), and 60.09% (intervention availability), outperforming baselines.
• Advantage over baselines: CALLM surpassed traditional and LLM baselines by leveraging curated context (similarity-matched peer cases and personal trajectories), demonstrating significant gains in inference accuracy.
• Experiments / evaluation: Evaluated on 24,183 diary entries from 407 cancer survivors using 5-fold grouped cross-validation. Metrics included balanced accuracy, ROC-AUC, and statistical tests for significance.
• Limitations and future work: Availability inference plateaued at ∼60%, suggesting the need for multimodal sensing. The sample skewed female and White, limiting generalizability. Future work should explore culturally informed modeling, richer contextual signals, and real-world JITAI deployment.

Summary

This study introduces CALLM, a context-aware framework for inferring affective states, regulation desire, and intervention availability from ultra-brief mobile diaries of cancer survivors. By grounding predictions in curated context—similarity-aligned peer cases and personal trajectories—CALLM achieves significant accuracy gains over baselines, with balanced accuracy reaching ∼73% for affective constructs. Findings inform the design of low-burden, confidence-gated digital interventions that respect user autonomy and situational context. Future work should address equity concerns, extend multimodal sensing, and explore real-world deployment scenarios for JITAIs.