julia-pfarr/HarmoniCA_depression

HarmoniCA — Depression Model

Part of the HarmoniCA collection for harmonising psychiatric questionnaire items across studies. This model assigns items from depression questionnaires to one of five theoretically motivated symptom dimensions.

Dimensions

ID	Label	Description
1	Mood & Affective Symptoms	dysphoria, sadness, hopelessness, feeling low
2	Cognitive & Self-Perception	guilt, worthlessness, self-criticism, pessimism, narratives of the past and future
3	Somatic & Vegetative Symptoms	sleep, appetite, fatigue, weight changes, physical symptoms
4	Activity & Interest Deficit	anhedonia, loss of interest, psychomotor signs, reduced functioning
5	Anxiety & Distress	irritability, agitation, worry, tension, nervousness

Items rated as not belonging to the depression construct are assigned dimension -1 (Does not fit).

Model Architecture

• Type: ContrastiveModel (contrastive fine-tuning of a sentence transformer)
• Base encoder: BAAI/bge-large-en-v1.5

Training Data

• Training items: 203 expert-labeled questionnaire items (164 original + 27 augmented + 12 expert-reviewed)
• Questionnaires: BDI, DSI, GDS, HADS, HAMD, MADRS, MDS-UPDRS, MFQ, NMSQuest, NMSS, NPI, PHQ-9, SCL-90, SDS, plus data augmentation
• Labels: Expert consensus from two rounds of surveys (psychiatrists/neurologists)

Evaluation

Evaluated on 126 held-out items from unseen questionnaires against expert labels.

Cohen's κ	Accuracy
0.746	79.4%

Cross-validation (leave-one-questionnaire-out, 14 folds): mean accuracy 86.9% ± 11.9%.

Limitations

• Trained and validated on Parkinson's disease cohort questionnaires; performance may differ in other populations.
• Dimension structure reflects theoretical consensus among a specific expert panel and may not match all theoretical frameworks.

Citation

in preparation