openenv.yaml

name: SepsisPilot
version: "1.0.0"
description: >
  Reinforcement learning environment for optimal sepsis treatment sequencing.
  An AI agent observes ICU patient vitals (MAP, lactate, WBC, temperature,
  heart rate, creatinine) and decides which antibiotic and vasopressor
  combination to administer each hour. The environment models realistic
  sepsis physiology including antibiotic resistance, multi-organ dysfunction,
  and haemodynamic instability.

  Built for the Meta PyTorch OpenEnv Hackathon 2026.

tags:
  - openenv
  - healthcare
  - reinforcement-learning
  - sepsis
  - icu
  - medical-ai

tasks:
  - name: mild_sepsis
    difficulty: easy
    max_steps: 24
    description: >
      Mild sepsis from gram-negative UTI. Broad-spectrum antibiotics are the
      correct choice. Vasopressors needed only if MAP < 65. Slow deterioration
      gives the agent time to learn.

  - name: septic_shock
    difficulty: medium
    max_steps: 48
    description: >
      Septic shock from gram-positive bacteraemia (MRSA). MAP critically low;
      vasopressors are mandatory immediately. Narrow-spectrum antibiotics
      (vancomycin) are optimal. Each delayed hour increases organ failure risk.

  - name: severe_mods
    difficulty: hard
    max_steps: 72
    description: >
      Severe sepsis with Multi-Organ Dysfunction Syndrome. Mixed drug-resistant
      infection requiring precise antibiotic sequencing: broad-spectrum first
      to cover gram-negative load, then switch to narrow-spectrum. High-dose
      vasopressors risk acute kidney injury. Antibiotic resistance accumulates
      with repeated suboptimal choices.

action_space:
  type: discrete
  n: 9
  actions:
    - id: 0
      name: no_treatment
      description: Watchful waiting — no intervention
    - id: 1
      name: broad_antibiotics
      description: Broad-spectrum antibiotics (piperacillin-tazobactam)
    - id: 2
      name: narrow_antibiotics
      description: Narrow-spectrum antibiotics (vancomycin)
    - id: 3
      name: low_vasopressor
      description: Low-dose norepinephrine (0.1 mcg/kg/min)
    - id: 4
      name: high_vasopressor
      description: High-dose norepinephrine (0.3 mcg/kg/min) — renal risk
    - id: 5
      name: broad_plus_low_vaso
      description: Broad-spectrum antibiotics + low-dose vasopressor
    - id: 6
      name: broad_plus_high_vaso
      description: Broad-spectrum antibiotics + high-dose vasopressor
    - id: 7
      name: narrow_plus_low_vaso
      description: Narrow-spectrum antibiotics + low-dose vasopressor
    - id: 8
      name: narrow_plus_high_vaso
      description: Narrow-spectrum antibiotics + high-dose vasopressor

observation_space:
  type: continuous
  shape: [9]
  fields:
    - name: map_mmhg
      description: Mean Arterial Pressure (mmHg). Sepsis goal ≥ 65.
      range: [20.0, 160.0]
    - name: lactate
      description: Serum lactate (mmol/L). Normal 0.5–2.0; crisis > 4.
      range: [0.1, 20.0]
    - name: wbc
      description: White blood cell count (k/uL). Normal 4–11.
      range: [0.5, 40.0]
    - name: temperature
      description: Core body temperature (°C). Sepsis > 38 or < 36.
      range: [33.0, 42.0]
    - name: heart_rate
      description: Heart rate (bpm). Sepsis > 90.
      range: [20.0, 170.0]
    - name: creatinine
      description: Serum creatinine (mg/dL). AKI marker.
      range: [0.3, 12.0]
    - name: sofa_score
      description: SOFA score (0–24). Organ failure composite.
      range: [0.0, 24.0]
    - name: resistance
      description: Antibiotic resistance index (0–1). Hard task only.
      range: [0.0, 1.0]
    - name: step_fraction
      description: Fraction of episode elapsed (step / max_steps).
      range: [0.0, 1.0]

reward:
  type: dense
  description: >
    Reward is provided at every timestep (dense shaping), not just at
    episode end. Each vital sign contributes incrementally:
      +0.35 per step if MAP ≥ 65
      +0.30 per step if lactate < 2.0
      +0.10 per step if WBC in normal range
      +0.08 per step if temperature normal
      −0.025 per step (time pressure)
      −8.0  on patient death
      +5.0  on full stabilisation (all vitals normal)
    Additionally, creatinine protection and resistance management
    contribute in medium/hard tasks.
  range: [-8.0, 5.775]

grader:
  type: composite
  output_range: [0.0, 1.0]
  components:
    - survival (mandatory prerequisite)
    - final vital sign normalisation
    - stabilisation speed
    - treatment appropriateness (correct antibiotic class)
    - organ protection (renal function)
    - resistance management (hard task)

infrastructure:
  runtime_limit_minutes: 20
  vcpu: 2
  memory_gb: 8
  port: 7860

api:
  reset:   POST /reset
  step:    POST /step
  state:   GET  /state
  grade:   GET  /grade
  tasks:   GET  /tasks
  health:  GET  /health

environment_variables:
  - name: API_BASE_URL
    description: LLM endpoint base URL
    default: "https://integrate.api.nvidia.com/v1"
  - name: MODEL_NAME
    description: LLM model identifier
    default: "nvidia/llama-3.1-nemotron-70b-instruct"
  - name: HF_TOKEN
    description: Hugging Face API token
  - name: OPENAI_API_KEY
    description: API key for the LLM endpoint (OpenAI-compatible)
  - name: ENV_BASE_URL
    description: SepsisPilot environment server URL
    default: "http://localhost:7860"