Integrative, and Scalable mental health phenotyping using a knowledge-graph-derived dual-metric framework

This repository contains all code, trained models, and reproducible workflows associated with the manuscript: “Integrative and Scalable Mental Health Phenotyping Using a Knowledge-Graph–Derived Dual-Metric Framework.” The project operationalizes two quantitative metrics , the Cognitive Attention Score (CAS) and the Cognition-to-Emotional-Response Ratio (C:ERR) into a structured knowledge graph (Ceekr-KG), and evaluates its structural validity, predictive performance, clinical benchmarking, and longitudinal sensitivity. The archive enables Full reconstruction of the Ceekr Knowledge Graph (151,288 triples; 460 nodes; 6 relation types) Reproduction of embedding-based validation analyses Re-training of baseline and hypertuned SimplE models Replication of relation-preserving shuffled graph controls Integration of Ceekr-KG with the Clinical Knowledge Graph (CKG) Reproduction of all manuscript figures (Figures 1–4) 🔧 Environment Setup # 1) Create a new conda environmentconda create -n ceekr_kg_env python=3.11 -y# 2) Activate the environmentconda activate ceekr_kg_env# 3) Install required Python packagesconda install -y numpy pandas tqdm# 4) Install PyTorch (CPU version)pip install torch=1.10.0+cu113# 5) Install DGLpip install dgl=0.5.2# 6) Install the modified DGL-KE package included in this repositorycd dgl-kepip install -e .cd .. 📁 Repository Structure 1️⃣ all_figures_R_code/ R scripts used to generate all manuscript figures. Figure_1_updated.RDistributional analysis, correlation heatmaps, Q–Q plots, and ordinal regression AIC comparisons for CAS and C:ERR. Figure_2_updated.RKnowledge graph construction visualization, embedding model comparison, shuffled graph validation, and cross-validation analyses. Figure_3_updated.RClinical Knowledge Graph (CKG) integration benchmarking and triple prediction score comparisons. Figure_4_updated.RLongitudinal cohort filtering, paired statistical analysis, CAS band transitions, and chi-square evaluation of user-reported agreement. These scripts reproduce all statistical analyses and visualizations reported in the manuscript. 2️⃣ CEEKR_KG_basemodel/ — Base Ceekr Knowledge Graph Model Contains trained embedding outputs for the standalone Ceekr Knowledge Graph. entities_final.dictMapping of entity names to integer IDs used during embedding training. relation.dictMapping of relation types to integer IDs. SimplE/SimplE_CEEKR_0/ CEEKR_SimplE_entity.npy — Learned entity embeddings CEEKR_SimplE_relation.npy — Learned relation embeddings config.json — Baseline training hyperparameters Represents the base SimplE model (Hits@1 ≈ 82.7%, MRR ≈ 0.90). 3️⃣ CEEKR_KG_Hypertuned/ Contains optimized SimplE model trained on Ceekr-KG with tuned hyperparameters(embedding size 128, gamma 21, learning rate 0.5, 2500 steps). CEEKR_SimplE_entity.npy CEEKR_SimplE_relation.npy config.json Represents the hypertuned model achieving: Hits@1 ≈ 96.9% Hits@3 ≈ 99.8% MRR ≈ 0.98 4️⃣ CEEKR_CKG_int/ — Integrated Ceekr–CKG Model Contains embedding outputs from the integration of Ceekr-KG into the Clinical Knowledge Graph (CKG). entities_final.dict relation.dict SimplE/SimplE_CEEKR_0/ CEEKR_SimplE_entity.npy CEEKR_SimplE_relation.npy config.json This integrated model supports benchmarking against HAM-A, HAM-D, and GAD-7 within a large biomedical knowledge graph environment. 5️⃣ CEEKR_KG_NB/ — Reproducible Construction & Training Workflows Contains Jupyter notebooks and training scripts enabling full reconstruction of the computational pipeline. CEEKR_KG.ipynbConstructs the Ceekr Knowledge Graph from the curated CAS–C:ERR–symptom matrix, generates triples, and prepares datasets for embedding training. CKG_DD_CEEKR.ipynbDemonstrates integration of Ceekr-KG into the Clinical Knowledge Graph (CKG), including entity alignment and merged graph preparation. CEEKR_KG_Simple.shShell script for training baseline and hypertuned SimplE models on standalone Ceekr-KG using DGL-KE. CEEKR_CKG_int_simple.shShell script for training SimplE on the integrated CKG–Ceekr graph. These files allow full reproducibility of all embedding-based validation results reported in the manuscript. 6️⃣ Zipped Archives The following compressed files are provided for convenience and direct download: CEEKR_KG_basemodel.zip CEEKR_KG_Hypertuned.zip CEEKR_CKG_int.zip CEEKR_KG_NB.zip