Synthica — ICU Trajectory Modeling

Research-grade machine learning system for ICU time-series modeling, targeting ICML-level rigor.

Objective

• Treatment-conditioned forecasting of physiological trajectories
• Counterfactual trajectory simulation under modified treatment sequences
• Hospital-invariant representation learning via disentangled latent dynamics
• Cross-domain generalization across ICUs / hospitals

Dataset

Download eicu_final_sequences_for_modeling.csv from Google Drive and place it at:

synthica-to-be-moved

ICML-Level ICU Trajectory Model

A research-grade machine learning system for ICU time-series modelling: treatment-conditioned forecasting, counterfactual trajectory simulation, and hospital-invariant representation learning.

---

Scientific Goal

"We learn invariant latent disease dynamics disentangled from hospital and treatment effects, enabling robust cross-hospital generalisation and counterfactual ICU trajectory simulation."

---

Dataset

Download the dataset from Google Drive and place it at:

data/eicu_final_sequences_for_modeling.csv

The file name must be exact. All data loading uses:

df = pd.read_csv("data/eicu_final_sequences_for_modeling.csv")

Download link: https://drive.google.com/file/d/1cjf5yjnQXpuMLJSyQlhQkn75g7Pi9GUK/view?usp=sharing

Note: The file name must remain exactly eicu_final_sequences_for_modeling.csv.

---

Quick Start

pip install torch scikit-learn pandas numpy
python src/train.py

Results are written to outputs/results.json.

Code Structure

src/
  data.py        — Dataset loading, schema inference, environment validation
  model.py       — DisentangledICUModel (s_t / e_t split, GRL adversaries)
  baselines.py   — All baseline models (CRN, RMSN, DCRN, G-Net, etc.)
  losses.py      — Reconstruction, outcome, adversarial, IRM, contrastive losses
  eval.py        — AUROC/AUPRC/ECE, counterfactual proxy eval, competitiveness flag
  train.py       — Main training loop, baseline tuning harness, ablation suite
  utils.py       — Seeding, early stopping, checkpointing
data/
  eicu_final_sequences_for_modeling.csv  (not committed — download separately)
outputs/
  results.json                — Full experiment results
  ece_main_model.json         — Calibration reliability data
  best_baseline_configs.json  — Best hyperparameter configs per baseline
  checkpoints/                — Model checkpoints

Key Features

1. Credible Baselines with Tuning

Hyperparameter search over hidden_dim ∈ {64,128,256}, lr ∈ {1e-3,3e-4,1e-4}, dropout ∈ {0,0.2,0.5}.
Best-of-k-trials config is logged and persisted.

2. Environment Validation

Environment/hospital column is inferred programmatically and validated:

• ≥ 3 unique environments required
• Non-trivial patient distribution per environment
• If invalid → hospital invariance mechanisms auto-disabled with prominent warning

3. Counterfactual Proxy Evaluation

Clearly labelled as proxy evaluation (not causal ground truth):

• Consistency test (same u_t → same rollout)
• Temporal intervention test (intervene at t=k vs t=k+Δ)
• Monotonicity sanity (skipped with logged reason if unknown)
• Stress tests (zero/max/flip treatments)

4. Calibration (Mandatory)

Expected Calibration Error (ECE) reported alongside AUROC/AUPRC/accuracy.
Reliability bin stats saved to disk for plotting.

5. Competitiveness Flagging

ΔAUROC = model - best_baseline logged after evaluation.
Prominent warning emitted if negative.

6. Key Ablation: No s_t

EtOnlyModel removes s_t entirely — uses only e_t — to validate that the invariant branch carries signal.

7. Invariant Latent Collapse Monitoring

s_t variance monitored during training; warning emitted if collapse detected.
Disentanglement check: performance degradation when e_t branch is zeroed.

Reproducibility

• Seed: 42 (deterministic splits and training)
• Early stopping on validation AUROC (patience=10)
• Best model checkpointed to outputs/checkpoints/main_model_best.pt pip install -r requirements.txt python src/train.py

The pipeline will run end-to-end automatically: no manual edits required.

Optional CLI flags:

```bash
python src/train.py --mode all           # train main model + baselines + ablations (default)
python src/train.py --mode main          # main model only (fastest)
python src/train.py --epochs 100         # override epoch count
python src/train.py --split_mode random  # random patient split instead of out-of-hospital
python src/train.py --output_dir runs/   # custom output directory
python src/train.py --no_ablations       # skip ablation studies
python src/train.py --no_baselines       # skip baselines

---

Architecture

                ┌────────────────────────────────────────┐
Input           │  GRUEncoder: (x_t ‖ u_t) → h_t        │
(x_t, u_t)  ───▶│      Linear split → s_t , e_t          │
                └────────────┬──────────────┬─────────────┘
                             │              │
                   s_t (invariant)      e_t (environment)
                             │              │
              ┌──────────────▼──┐   ┌───────▼──────────────┐
              │ InvariantDyn.   │   │  EnvironmentDyn.     │
              │ s_{t+1}=GRU(s_t)│   │  e_{t+1}=GRU(e_t,u_t)│
              └──────────────┬──┘   └───────┬──────────────┘
                             │              │
                        ┌────▼──────────────▼────┐
                        │  Decoder: MLP(s‖e)→x̂  │
                        └─────────────────────────┘

Invariance:  HospitalAdversary(GRL → s_t)  → hospital label
             TreatmentAdversary(GRL → s_t) → treatment label
             ContrastiveHead(s_t)           → NT-Xent alignment
Outcome:     OutcomeHead(s_T) → mortality logit

---

Loss Function

L = L_reconstruction
  + λ₁ · L_outcome
  + λ₂ · L_hospital_adversary
  + λ₃ · L_treatment_adversary
  + λ₄ · L_contrastive          (optional)
  + λ₅ · L_IRM                  (optional)

Discriminators are trained in a separate optimiser step (no gradient reversal needed for their update); the encoder is trained with gradient reversal to fool them.

---

Baselines

Name	Category	Reference
CRN	Counterfactual	Bica et al., NeurIPS 2020
G-Net	Counterfactual	Li et al., 2021
RMSN	Counterfactual	Lim et al., 2018
DCRN	Counterfactual	—
CausalTransformer	Counterfactual	Melnychuk et al., NeurIPS 2022
GTransformer	Counterfactual	Li et al., 2021
MambaCDSP	Counterfactual	Simplified SSM
ERM	Hospital invariance	Standard supervised GRU
DANN	Hospital invariance	Ganin et al., JMLR 2016
DomainConfusion	Hospital invariance	Tzeng et al., 2015
IRM	Hospital invariance	Arjovsky et al., 2019

---

Experiments

1. In-distribution — random 70/10/20 patient split (seed=42)
2. Out-of-hospital — train on N-k hospitals, test on held-out k hospitals
3. Counterfactual — intervene on u_t, measure trajectory divergence & outcome sensitivity
4. Ablation — remove adversarial heads, e_t branch, contrastive loss

---

Metrics

• AUROC (primary)
• AUPRC
• Accuracy
• Expected Calibration Error (ECE)
• Domain shift degradation (in-dist vs. OOH AUROC drop)
• Counterfactual sensitivity score

---

Repository Structure

src/
  data.py        — Data loading, schema discovery, sequence construction, splits
  model.py       — DisentangledICUModel (main architecture)
  baselines.py   — 11 SOTA baseline implementations
  train.py       — Training loop, early stopping, checkpointing, ablations
  eval.py        — All evaluation metrics and report generation
  losses.py      — Reconstruction, outcome, adversarial, contrastive, IRM losses
  utils.py       — GradientReversalLayer, logging, seeding, metric helpers
data/
  README.md      — Dataset download instructions
requirements.txt

---

Outputs

After training, the following outputs are written to outputs/ (default):

outputs/
  config.json           — Resolved configuration
  results.json          — Full evaluation report (all models, all experiments)
  train.log             — Training log
  checkpoints/
    best_main.pt        — Best main model checkpoint
    best_<baseline>.pt  — Per-baseline checkpoints
    ablation_*.pt       — Ablation checkpoints