Source-linked AI summary
Cross-Scale Transfer Learning for Depression Severity Prediction: From PHQ-8 to HAMD-17 Across Languages and Clinical Paradigms
Wenjie Feng, Sahba Zojaji, Satoshi Nakamura
TL;DR
Continuous whole-scale depression-severity prediction across assessment scales remains unaddressed in transcript research. This study evaluates sequential transfer from English PHQ-8 to Chinese HAMD-17, with the best point estimates on both backbones, while treating the evidence as exploratory.
Problem
Continuous whole-scale severity prediction from transcripts across depression scales remains unaddressed, despite the need for more efficient depression assessment.
Source: §1 · p. 1Method
Results
Sequential transfer had the best point estimates among configurations: 4.96 MAE and 0.36 macro-F1 at 0.6B, and 4.38 MAE and 0.46 macro-F1 at 1.7B, with gaps within per-run variance.
Source: §4 · p. 4Takeaways & Limitations
The study proposes related, differently scaled data as a strategy for augmenting small clinical datasets, while noting that the ablations do not isolate alignment as the sole driver.
Source: §1 · p. 2Evidence is limited to 100 inpatients at one hospital and a single source–target pair; scale and language remain confounded, and observed gaps are not significance claims.
Source: §5 · p. 4
Abstract
from arXiv · showhide
This work addresses continuous depression-severity score prediction from clinical interview transcripts under data scarcity. We propose a sequential low-rank adaptation (LoRA) 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 adapter then initializes fine-tuning on the Chinese PDCH dataset (100 real clinical consultations, HAMD-17), where a reinitialised, scale-specific head predicts the clinician-assigned score. All configurations use patient-level stratified 5-fold, 2-repeat cross-validation. On the data-scarce HAMD-17 target, the sequential protocol attains the best point-estimate MAE , RMSE, and macro-$F_1$ on both 0.6B and 1.7B backbones, outperforming target-only training and non-LLM baselines---4.96/6.59/0.36 with Qwen3-0.6B and 4.38/5.62/0.46 with Qwen3-1.7B. Ablations suggest that correctly aligned source supervision gives the best point estimates (unsupervised exposure and shuffled-label controls also show partial gains), that native-Chinese target input outperforms machine-translated English input, and that the reversed order yields no clear gain within run-to-run variance. The study is an exploratory, single-site internal evaluation: it does not establish screening or diagnostic utility, nor separately identify the contribution of the scale, language, or paradigm shifts. To our knowledge, no prior study evaluates this specific DAIC-WOZ-to-PDCH sequential transfer setting.
1. INTRODUCTION
This study asks whether depression-severity representations transfer across related but nonequivalent scales, languages, and interview settings. It proposes a sequential transfer evaluation while acknowledging that these shifts cannot be disentangled in this experiment.
1. INTRODUCTION
- DAIC-WOZ provides 189 English avatar-mediated sessions scored with PHQ-8, whereas PDCH provides 100 Chinese real-clinical consultations scored with HAMD-17.PDCH use has so far been preliminary, mostly item-level or binary-cutoff classification.Source: §1 · p. 1
- Depression severity prediction from transcripts across PHQ-8 and HAMD-17 remains unaddressed, while existing work largely targets binary classification or a single scale and setting.The study frames the task as transfer between related but nonequivalent instruments, whose partial agreement bounds transferable severity signal.Source: §1 · p. 1
- The study evaluates one combined shift in scale, language, and collection paradigm, so it cannot separately identify each shift’s contribution.It is presented as cross-scale transfer rather than transfer of a shared severity scale.Source: §1 · p. 1
- The proposed pipeline adapts Qwen3 with LoRA and a bounded regression head on DAIC-WOZ PHQ-8, then transfers the adapter to PDCH HAMD-17 with a reinitialised head.The target dataset contains only 100 sessions, motivating use of the larger, differently scaled source dataset.Source: §1 · p. 2
- The work contributes a cross-scale continuous-regression study, controlled transfer ablations, and a reproducible pipeline evaluated with patient-level stratified 5-fold, 2-repeat group cross-validation.Correctly aligned source supervision gives the best point estimates, but the controls do not isolate alignment as the sole driver.Source: §1 · p. 2
- The PDCH input is ASR-derived from tone-converted recordings, so the study focuses on text and leaves acoustic follow-up, including ASR-error analysis, for future work.The altered acoustic channel is not comparable to raw audio.Source: §1 · p. 2
2. METHODS
The method reads a dialogue representation from a frozen Qwen3 backbone adapted with LoRA and maps it to a bounded continuous severity score. The source adapter is transferred to the target task, while the scale-specific score head is re-estimated.
Architecture
- A frozen Qwen3-0.6B or 1.7B backbone is adapted through rank-8 LoRA in attention and MLP projections, alongside a trainable linear regression head.The original backbone parameters remain frozen; only the adapters and head are updated.Source: §2 · p. 2
Input construction and readout
- The transcript is placed in a language-matched analysis prompt, and the final-layer hidden state at the last non-padding token supplies the dialogue representation.The prompt asks for symptom and severity analysis rather than a numeric answer.Source: §2 · p. 2
Score prediction
- The regression head produces a continuous score in (0, S) by sigmoid-rescaling its output, where S is the scale maximum.S is 24 for PHQ-8 and 52 for HAMD-17; no token is decoded to produce the score.Source: §2 · p. 2
Training objective
- Training uses smooth-L1 loss with β=1 between predicted and ground-truth scores, with the LM head bypassed and updates driven by regression loss.AdamW, a 10^-4 learning rate, gradient clipping, and validation-MAE early stopping are also specified.Source: §2 · p. 2
Evaluation
- Evaluation reports MAE and RMSE plus threshold-based severity-band accuracy and macro-F1; the bands are reporting conventions, not diagnostic classifications.The HAMD-17 scheme uses minimal 0–7, mild 8–16, moderate 17–23, and severe 24–52 thresholds.Source: §2 · p. 2
Two-stage transfer
- Two-stage transfer trains a fresh LoRA and head on a random 90% DAIC-WOZ session-level subsample (170/189; PHQ-8, seed 42), then loads trainable LoRA into PDCH (HAMD-17; ceiling S=52) and reinitializes the head.Source: §2 · p. 2
3. EXPERIMENTS
The experiments compare sequential DAIC-WOZ-to-PDCH transfer with target-only training and controls that test source supervision, language, transcript content, and transfer order. On PDCH, Mix has the best point estimates for MAE, RMSE, and macro-F1 on both backbones.
3.2. Baselines and Ablations
- The ablation design contrasts target-only training, aligned and shuffled source labels, unsupervised DAIC exposure, machine-translated target text, reverse transfer, and speaker-only transcripts.The comparisons are organized around whether transfer helps, whether DAIC contributes, whether language is a confound, and where transcript signal resides.Source: §3.2 · p. 3
- The study also uses mean, median, TF-IDF plus Ridge, and zero-shot Qwen3 as reference baselines; zero-shot prompting produced scores far from ground truth.The zero-shot MAE was 14.01 at 0.6B and 15.47 at 1.7B, compared with 5.98 for the non-LLM TF-IDF+Ridge floor.Source: §3.2 · p. 3, §3.3 · p. 4
3.3. Main Results
- Across configurations, Qwen3-0.6B recorded 4.96 MAE, 6.59 RMSE, and 0.36 macro-F1; Qwen3-1.7B recorded 4.38 MAE, 5.62 RMSE, and 0.46 macro-F1.Source: §3.3 · p. 4
- Aligned source supervision gave the best point estimates on both backbones, while unsupervised and shuffled-label controls also showed partial gains and reverse-order results were inconclusive within run-to-run variance.The controls do not establish that alignment alone drives the gain; the shuffled-label permutation was fixed and controls were not compute-matched.Source: §3.3 · p. 4
- Native-Chinese target input outperformed machine-translated English input on both backbones, though translation effects and asymmetric language capacity may also contribute.At 0.6B, Mix versus All-English was 4.96 versus 5.99 MAE and 0.36 versus 0.24 macro-F1; at 1.7B, it was 4.38 versus 5.09 MAE and 0.46 versus 0.41 macro-F1.Source: §3.3 · p. 4
- The full dialogue was the best input at both sizes; removing patient turns cost more MAE than removing therapist turns, but the comparison does not isolate causal speaker-specific signal.Removing patient turns cost 0.82 MAE at 0.6B and 1.16 at 1.7B; removing therapist turns cost 0.33 and 0.08, respectively.Source: §3.3 · p. 4
4. CONCLUSION
Sequential LoRA transfer from DAIC-WOZ to PDCH produced the best point estimates among tested configurations on the data-scarce HAMD-17 target. Results are preliminary and do not establish screening or diagnostic utility.
4. CONCLUSION
- Mix achieved 4.96 MAE and 0.36 macro-F1 at 0.6B, and 4.38 MAE and 0.46 macro-F1 at 1.7B, versus 5.50/0.31 and 4.55/0.36 without transfer.The reported gaps are within per-run variance and are point estimates, not validated improvements.Source: §4 · p. 4
- Correctly aligned source supervision gave the best point estimates, while unsupervised exposure and shuffled-label adaptation showed partial, size-inconsistent gains.Next-token training helped at 0.6B but degraded at 1.7B; shuffled labels hurt at 0.6B and approached Mix at 1.7B.Source: §4 · p. 4
- Native-Chinese target input outperformed machine-translated English, while reversed-order transfer showed no clear gain within run-to-run variance.The study is an exploratory, single-site proof of concept and does not establish screening or diagnostic utility.Source: §4 · p. 4
5. LIMITATIONS
The evaluation is constrained by a small, single-site target cohort, confounded shifts, and transcript and tuning conditions that limit generalisation and interpretation.
5. LIMITATIONS
- PDCH comprises 100 inpatients from one hospital, and observed MAE gaps of 0.17–0.54 are of the same order as per-run variance, so results are trends rather than significance claims.The skewed band distribution makes band-level metrics noisy and CV stratification marginal.Source: §5 · p. 4
- Scale and language remain confounded because each scale appears in only one language, and one source–target pair limits generalisability.The All-English ablation only partially de-confounds the axes.Source: §5 · p. 4
- DAIC-WOZ avatar-mediated English dialogue and PDCH Chinese inpatient consultations differ in paradigm, while ASR transcripts and privacy-processing artefacts may interact with scale- or language-specific effects.Altered audio precludes acoustic follow-up; Stage-1 hyperparameters also carry over to Stage 2 without per-stage retuning.Source: §5 · p. 4
6. COMPLIANCE WITH ETHICAL STANDARDS
The study retrospectively analysed previously collected, de-identified human-subject data under their existing ethical approvals, with no new collection or re-identification.
6. COMPLIANCE WITH ETHICAL STANDARDS
- The work was a retrospective analysis of previously collected human-subject data, with no new data collection or re-identification.Source: §6 · p. 5
- PDCH data comprised 100 Chinese HAMD-17 clinical consultations made available in open access by the Chinese Academy of Sciences’ Science Data Bank.The dataset is described in.Source: §6 · p. 5
- DAIC-WOZ data comprised 189 English PHQ-8 sessions obtained from USC’s Information Sciences Institute under its standard data license.Source: §6 · p. 5