NeuroTrack-LPF

# 🧠 Latent Psychodynamic Flow (LPF) A probabilistic deep learning framework for modeling psychiatric symptom progression using multimodal time series data. ## 📝 Overview Latent Psychodynamic Flow (LPF) is an AI-powered framework designed to model the nonlinear, temporal dynamics of psychiatric disorders. It combines **probabilistic temporal modeling**, **attention-guided encoding**, and **domain-aware variational inference** to predict symptom severity and diagnostic trajectories with interpretability and clinical relevance. ## 📌 Key Features - **Flow-based latent transition modeling** to capture complex, nonlinear psychiatric state evolution.- **Attention-guided variational inference** for interpretable latent trajectory learning.- **Multimodal fusion** integrating structured and unstructured clinical data (e.g., self-reports, physiological signals, neuroimaging).- **Domain-Aware Variational Synchronization (DVS)** for robust training under sparse and irregular observations.- **Clinical priors and structural constraints** enhance prediction stability and interpretability. ## 🧠 Architecture ### 1. Latent Psychodynamic Flow (LPF) - **Encoder**: Attention-guided encoder maps raw input to latent variables via multi-head attention and RNN.- **Latent Transition**: Flow-based transition dynamics model psychiatric state evolution over time.- **Decoder**: Maps latent state to clinical outcomes via hierarchical observation-outcome modeling. _See architecture diagram on page 5 of the paper._ ### 2. Domain-Aware Variational Synchronization (DVS) - Synchronizes latent representations across irregular time intervals.- Embeds clinical priors using personalized variational distributions.- Enforces **latent structural consistency** across patient cohorts. _Detailed in page 6–8 with schematic illustrations._ ## 🧪 Datasets We evaluated LPF on four prominent medical imaging datasets: - **ChestX-ray14**: Multi-label disease classification (NIH).- **LUNA16**: Lung nodule detection (CT scans).- **BraTS**: Brain tumor segmentation (MRI).- **CAMELYON16**: Histopathology images for metastasis detection. ## 📈 Performance LPF consistently outperformed SOTA models such as Informer, Autoformer, and PatchTST across multiple metrics: | Dataset | Accuracy | RMSE | R² Score ||---------------|----------|-------|----------|| ChestX-ray14 | 90.13% | 0.164 | 0.851 || LUNA16 | 89.47% | 0.169 | 0.842 || BraTS | 89.61% | 0.188 | 0.823 || CAMELYON16 | 88.97% | 0.193 | 0.812 | More benchmarks in Tables 1 and 2 (pages 11–12). ## 🔍 Ablation Study We ablated key components: - **w/o Flow Transition** → major drop in R²- **w/o HOOM** → less precise outcome mapping- **w/o DVS** → reduced robustness under sparse sampling 📄 See Tables 3 and 4 for quantitative results. ## ⚙️ Installation ```bashgit clone https://github.com/your-username/LPF-AI-Psychiatry.gitcd LPF-AI-Psychiatrypip install -r requirements.txt