Bed Admissions Manager Optimization

NSGA-II multi-objective optimizer for hospital bed assignments. Returns a Pareto front of non-dominated solutions across three competing clinical objectives; the charge nurse selects the operating point that matches shift priorities.

---

Problem

Assigning patients to beds requires balancing goals that conflict:

• Placing every patient near their clinical team (f1) pressures scarce anchor-ward beds.
• Filling primary beds before overflow (f2) may force cross-floor placements.
• Minimizing patient transfers (f3) constrains how much the plan can improve the above.

A weighted-sum formulation requires weights chosen in advance and produces one answer. NSGA-II surfaces the full trade-off space in a single run. The recommended solution is the knee point -- minimum L2 distance to the utopian point after min-max normalization, i.e. compromise programming with equal implicit weights.

---

Objectives

All three are minimized (lower = better).

ID	Dashboard Label	Definition
f1	Clinical Home Alignment	Sum of radius penalties between each patient's assigned bed and their specialty anchor ward. Penalty ladder: same ward = 1 pt, different wing = 5 pt, different floor = 10 pt
f2	Capacity Throughput	Count of primary (non-overflow) beds left unoccupied. Minimizing this maximizes ED intake capacity
f3	Nursing Workload Protection	CMI-weighted cost of patient transfers. Each move is scaled by Case Mix Index to reflect clinical complexity

Hard constraint: total moves <= transfer cap (default 5; adjustable at runtime).

---

Repository Layout

config/
  settings.yaml          Penalty weights, displacement cost, transfer cap default
  hospital_map.yaml      Ward topology (floor, wing, specialty anchor)

src/
  engine/
    models.py            Patient, Bed, Wing, DisplacementStatus dataclasses
    problem.py           WardAssignmentProblem (pymoo ElementwiseProblem)
    solver.py            NoDuplicateBedRepair, SolverConfig, run_optimization
    directives.py        DirectiveGenerator, ActionPlan, find_knee_point
  utils/
    mock_data.py         Synthetic 20-patient / 30-bed scenario
    exporters.py         ActionPlan -> JSON / Markdown
    sanitizer.py         HMAC pseudonymization + HIPAA Safe Harbor masking
  ui/
    dashboard.py         Streamlit command center

brand_style.css          mm.farina brand identity (Anthracite / Teal / Inter)
remotion-video/          45-50s NSGA-II explainer video (React + Remotion + Three.js)

---

Setup

pip install -r requirements.txt

Python 3.11+ required.

Run the dashboard:

streamlit run src/ui/dashboard.py

Remotion video (optional):

cd remotion-video
npm install
npm start          # Remotion Studio with hot reload
npm run build      # Render to out/ClinicalFlow.mp4

---

Configuration

Edit config/settings.yaml to tune penalties and solver defaults without touching code.

optimization:
  penalties:
    same_floor: 1      # home ward penalty (f1 minimum unit)
    diff_floor: 5      # same floor, different wing
    diff_wing: 10      # different floor
  displacement:
    cost_per_cmi_point: 2.5   # f3 weight per CMI unit
    max_transfers_default: 5

Edit config/hospital_map.yaml to change ward topology:

wards:
  - {id: "4A", specialty: "Cardiology", floor: 4, wing: "North"}
  ...

Runtime overrides are available from the dashboard sidebar (aggression multiplier, transfer cap slider, ED Surge Mode toggle).

---

Architecture Notes

Encoding: Integer vector x of length n_patients; x[i] is the index into the beds list for patient i. Allows overflow reuse and unassigned slots, which a permutation encoding cannot.

Repair operator: NoDuplicateBedRepair enforces the no-duplicate-bed hard constraint after each crossover/mutation step using vectorized np.unique and np.setdiff1d.

Caching: @st.cache_data on _run_cached. Non-hashable arguments (list[Patient], list[Bed]) are prefixed _ to opt out of hashing; primitive parameters form the effective cache key.

Pre-assignment mode: When confirmed discharges are known, the longest-LOS assigned patients are removed from the optimization. Their beds are treated as available, enabling pre-assignment of incoming admissions.

Directive categories:

• MOVE -- transfer improves alignment by more than move_cmi_threshold * cmi
• HOLD -- vacant anchor-ward bed; specialty demand is displaced elsewhere; reserve it
• ALERT -- no anchor-ward bed available; displacement is structural, not resolvable by reshuffling

---

Solver Scaling

Patients	pop_size	n_gen	Approx. runtime
20	100	100	< 5 s
50	150	150	~20 s
100	200	200	~90 s
200+	300	300	cache or async

---

Brand

Anthracite #36454F / Teal #008080 / Inter (body) / JetBrains Mono (code). Defined in brand_style.css and mirrored in src/ui/dashboard.py and remotion-video/src/constants.ts.