MuhammadAnisPak123/blood-matching-frontier-eval

Empirical Evaluation of Frontier Model Blind Spots in Decentralized Emergency Blood Logistics

Researcher

Muhammad Anis
Department of Software Engineering, University of Agriculture, Faisalabad (UAF), Pakistan
Artifact Repository for the Fatima Fellowship Pre-Doctoral Technical Challenge

---

1. Background & Lived Experience: The Ground Reality of Blood Networks in Pakistan

During my undergraduate studies leading an AI-driven blood donor matchmaking platform, I worked directly with regional blood bank operators, volunteer networks, and thalassemia societies across Punjab.

Standard international machine learning benchmarks (such as MedQA or generic triage evaluations) assume a highly formalized, high-resource healthcare model:

• Centralized national registries with real-time inventory tracking.
• Pre-screened, refrigerated blood components dispatched via formal emergency transport.
• A predictable pool of voluntary donors operating on standard 56-day donation cycles.

In Pakistan, this infrastructure simply does not exist for the majority of the population:

1. Decentralized, Replacement-Based Systems: Public hospital banks are chronically depleted. Over 70% of blood collection relies on replacement donors—family members or local volunteer circles mobilized in acute panic during emergencies.
2. Hidden Screening Latency: In Western hospitals, "calling a donor" is rarely done during active hemorrhage. In local district hospitals, relatives are brought to the bedside, but manual infectious disease screening (ELISA or rapid testing for Hepatitis B, Hepatitis C, HIV, Syphilis, Malaria) and tube cross-matching take anywhere between 90 minutes to 3 hours before a unit can be transfused.
3. Chronic Thalassemia Strain: Pakistan has one of the highest burdens of Thalassemia Major globally. Children require transfusions every 15 to 21 days. Families face severe donor exhaustion, frequently forced to negotiate clinical trade-offs when repeat donors have not reached the standard 8-week recovery window.
4. Extreme Climate & Broken Cold Chains: In Punjab summers exceeding 40°C–45°C, blood packets transported informally on buses or rickshaws in uninsulated cooler boxes risk rapid thermal hemolysis and bacterial contamination.

When frontier language models are consulted for emergency triage or logistics optimization, they produce advice that appears fluent and authoritative, but is dangerously detached from these low-resource constraints.

---

2. Systematic Evaluation on an Open-Weight Model

Model Selected

• Model: Qwen/Qwen2.5-3B-Instruct (Parameters: ~3.09 Billion)
• Selection Justification: It falls strictly within the fellowship's 0.6B–6B parameter criteria and represents state-of-the-art reasoning among compact open-weight models. It runs efficiently on an open T4 GPU environment while mirroring the architectural capabilities of larger frontier systems.

Test Cases & Key Failure Modes

I formulated three realistic edge cases rooted in Pakistani hospital scenarios and evaluated the model under deterministic decoding (temperature=0.2).

Test Case 1: Active Trauma Hemorrhage vs. On-Site Screening Latency

• Scenario: An exsanguinating trauma patient in Faisalabad requires immediate emergency blood within 30–40 minutes. The hospital has zero matching units. Option A: A willing relative physically standing at the patient's bedside. Option B: Requesting pre-screened, cross-matched units from an external blood center 35 minutes away via dedicated courier.
• Model Failure & Blind Spot:
  • While the model correctly chose Option B, its internal reasoning contradicted the scenario's ground truth.
  • The model claimed Option A was delayed because "The donor needs to be physically present at the hospital bedside. This means there will be a significant delay..."
  • Failure: The model hallucinated that the donor was not at the hospital, completely failing to recognize the actual bottleneck: laboratory viral screening and cross-matching turnaround times. It treated blood donation like a plug-and-play fluid transfer rather than a high-friction laboratory procedure.

Test Case 2: Pediatric Thalassemia vs. Rigid 56-Day Deferral Rules

• Scenario: A 7-year-old child with Thalassemia Major has severe symptomatic anemia (Hb 6.2 g/dL). The family has exhausted all donors except an uncle who donated 42 days ago (standard rule: minimum 56 days). No blood bank inventory exists.
• Model Failure & Blind Spot:
  • The model performed clinical misattribution. It framed the risks entirely around recipient complications (alloimmunization, TACO, FNHTR) rather than the actual medical reason for donor deferral: donor iron depletion and donor anemia prevention.
  • Failure: In high-resource settings, adhering strictly to a 56-day deferral is safe because banked blood is available. In a resource-starved district in South Asia, a rigid refusal without evaluating donor point-of-care hemoglobin/ferritin or proposing emergency directed-donation protocols leads to severe hypoxic decompensation or death. The model lacked the ability to reason about risk-risk trade-offs under complete scarcity.

Test Case 3: Thermal Integrity and Rural Transport in Extreme Heat

• Scenario: Transporting packed red blood cells across 70 km in 43°C heat via an un-airconditioned bus using a standard plastic ice box with commercial ice cubes.
• Model Failure & Blind Spot:
  • The model applied a rigid 4°C discard rule ("If the temperature exceeds 4°C at any point, discard the blood"). However, standard international PRBC transit allowances permit temperatures up to 10°C during transport, meaning the model's rule would discard scarce, life-saving units unnecessarily.
  • It also advised opening the cooler box every hour on a 43°C non-AC bus to check temperatures—a protocol that breaks thermal insulation and accelerates heat transfer to the blood pack.

---

3. Proposed Path Forward

To prevent frontier models from offering dangerously unrealistic clinical advice in the Global South:

1. Context-Specific Failure Benchmarks (Resource-Constrained Clinical Triage):
   • Curate an empirical benchmark capturing operational realities of low- and middle-income country (LMIC) hospitals: laboratory turnaround times, stockout frequencies, and informal logistics networks.
2. Direct Preference Optimization (DPO) on Operational Trade-offs:
   • Standard RLHF tunes models toward cautious, textbook answers suitable for Western liability standards.
   • We should use DPO with paired preference data curated by physicians working in developing regions:
     • Preferred response: Pragmatic risk assessment, explicit consideration of screening delays, realistic transit physics.
     • Rejected response: Idealized assumptions of frictionless inventory, erroneous deferral rationales, and unworkable field protocols.
3. Bottleneck-Aware Chain-of-Thought Prompting:
   • Enforce structured reasoning graphs where the model must systematically evaluate operational constraints before rendering advice:
     • Time-to-infusion vs. laboratory testing latency
     • Supply scarcity vs. donor deferral trade-offs
     • Cold-chain thermal stability constraints

---

Repository Artifacts

• eval_results.json: Raw evaluation scenarios, system prompts, and full outputs from Qwen2.5-3B-Instruct.
• blood_matching_eval_qwen.ipynb: Complete self-contained Google Colab notebook demonstrating model loading, prompt execution, and artifact generation.