AI Research

Cross-Scale Transfer Learning for Depression Severity Prediction: From PHQ-8 to HAMD-17 Across Languages and Clinical Paradigms | AI Learning Platform

BenchmarkMultimodalFine-tuningCost Optimization

Cross-Scale Transfer Learning for Depression Severity Prediction: From PHQ-8 to HAMD-17 Across Languages and Clinical Paradigms

Wenjie Feng, Sahba Zojaji, Satoshi Nakamura1 min readarXivSep 23, 2026

Key Takeaway for Students

This work addresses continuous depression-severity score prediction from clinical interview transcripts under data scarcity. We propose a sequential low-rank adaptation (efficient fine-tuning) protocol for cross-scale transfer: a Qwen3 backbone with a bounded regression head is first fine-tuned on the English DAIC-WOZ dataset (189 avatar-mediated sessions, PHQ-8), and the adapt

outperforming target-only training and non-LLM baselines---4outperforms machine-translated English input, and that the reversed or

Reading in Plain English mode — toggle below to switch

This work demonstrates an effective cross-scale and cross-lingual transfer learning approach for clinical NLP, particularly useful in data-scarce medical domains. ML engineers can leverage these techniques to build robust models for sensitive applications with limited annotated data, optimizing resource use and improving model generalization across different measurement scales and languages.

Read original on arXiv

Career Insight

high demand

Why it matters

Skills to learn

Implement sequential LoRA fine-tuning for transfer learningDevelop bounded regression heads for continuous score predictionApply cross-validation strategies for robust model evaluation on small datasetsUtilize PEFT library for efficient adapter-based fine-tuningEvaluate model performance using MAE, RMSE, and macro-F1 for regression tasks

Interview questions

Describe the sequential LoRA protocol for cross-scale transfer learning and explain why it's effective for data-scarce scenarios.
How would you adapt this cross-scale transfer learning approach if the target dataset had a different modality, such as audio or video, in addition to transcripts?
Design a system architecture for deploying this depression severity prediction model in a clinical setting, considering data privacy, real-time inference, and continuous model improvement with new clinical data.

Relevant for ~20 months · generated by the enrichment pipeline