OrangeOfficial/Ahsan_FatimaFellowshipApplication

Whole-Consultation Information Preservation in Code-Mixed Clinical Encounters

Fatima Fellowship Technical Challenge: Blind Spots of Frontier Models
Evaluated Model: Qwen/Qwen3.5-4B (4 Billion parameters, open-weight on Hugging Face)
Repository / Bucket: Public Artifacts & Evaluation Suite

---

Abstract

In overburdened outpatient clinics across Pakistan, doctors see 50 to 70 patients in a single morning with zero digital records. Ambient AI scribes could transform this workflow, but modern models are built around Western, English clinical standards. In this project, I evaluated an open-weight 4B model (Qwen/Qwen3.5-4B) on outpatient visits conducted in colloquial code-mixed Urdu/Punjabi-English compared to standard English.

Evaluating 210 clinical facts across 42 generated notes revealed three critical blind spots:

1. A 12.4% language penalty: Fact retention dropped from 90.5% in English to 78.1% in code-mixed dialogue simply because the patient spoke in their native language.
2. Loss of the patient's voice: The model dropped colloquial symptom phrases (like morning nausea) and mixed up who said what (putting a patient's home remedy of 7Up into the doctor's active treatment plan).
3. Medication collapse (47.6% retention): More than half of prescriptions were dropped or distorted. Without a glossary, the model didn't recognize local brands and guessed dangerous Western lookalikes—turning an anti-ulcer pill (Risek 40mg) into a drug that actively causes severe ulcers (Risedronate 40mg).

To fix this, we need a practical two-step approach:

• Immediate zero-cost fixes: Giving the model a simple local drug lookup list in the prompt stopped brand hallucinations completely, while separating patient notes from doctor orders stops the model from mixing up speakers.
• Better translation of local idioms: To stop models from dropping cultural phrases like "seene vich khich" (pulling chest tightness), small models should be fine-tuned with Direct Preference Optimization (DPO) so they learn to preserve descriptive physical feelings rather than dropping them.

---

Question 1: The Problem and Lived Experience

Recently I got the chance to work with my first overseas healthcare client, where i got first look at automated clinical documentation, electronic health records, and strict HIPAA compliance. It was impressive.

But when someone walks into an outpatient clinic in any city of Pakistan, the contrast is jarring. Doctors there see 50 to 70 patients in a single morning with no digital systems at all. A whole consultation ends up as a hurried, barely legible prescription slip (parchee) with a brand name and a scrawled dose. When you come back for a follow-up, or with a different problem, the doctor looks at a blank register and asks, "What medicine did I give you last time?" Nothing was recorded about your symptoms, the home remedies you tried, or the non-drug advice you were given.

An AI documentation assistant could really help these clinicians, but Western models struggle in this setting. Our consultations flow between Urdu, Punjabi, and English, with local brand names (Risek, Gabica) and colloquial symptom phrases ("seene vich khich", "talway jalna"). Before anyone talks about digital records, there’s a basic safety question to answer: when an AI documents a code-mixed consultation, does it keep the key clinical facts, including the patient’s symptoms and history? Or does it erase the patient’s voice, mix up who said what, and invent the wrong medicines?

---

Question 2: Methodology and Findings

Setup

I benchmarked an open-weight model, Qwen/Qwen3.5-4B, on 7 synthetic outpatient consultations. My brother, a medical student, wrote them with AI assistance. They cover GERD, diabetic neuropathy, hypertension, gastroenteritis, dysuria, diarrhea, and vertigo.

Each case existed in two versions with identical clinical facts:

1. Code-mixed Urdu/Punjabi-English (natural clinic conversational register)
2. Standard English (control condition)

Tested across three prompt conditions:

• Faithful: Record only what was said, in a structured SOAP note.
• Standardize: Convert brand names to generic names (INN) and colloquial terms to standard medical language.
• Glossary: The same as Standardize, plus a 10-word local clinic glossary.

I scored 210 tagged clinical facts across 42 generated notes using a frozen five-way rubric (Preserved, Omitted, Distorted, Attribution Error, Hallucinated), with Wilson 95% confidence intervals.

---

What I Found

1. A 12.4-point language penalty:
   Standard English retained 90.5% of facts (95/105; CI 83.4–94.7%). Code-mixed retained 78.1% (82/105; CI 69.3–84.9%). The exact same clinical case lost more than 12% of its information just because the conversation happened in the patient’s own language.

2. The patient’s voice erodes:

   • Dropped symptoms: In Scenario 1, the patient’s morning nausea and vomiting ("subah uthte hi ulti jaisi lagti hai") vanished from the code-mixed notes. In English it was kept every time.
   • Dropped context: In Scenario 2, the patient said their sugar spiked because they’d run out of medicine for 4 days ("dawai beech mein choot gayi thi"). The code-mixed notes left this out, which could lead a reader to wrongly conclude the drug wasn’t working.
   • Wrong speaker: In Scenario 4, the patient had taken 7Up with lemon at home. One code-mixed note put it in the doctor’s treatment plan.

3. Medication instructions collapsed:
   Vitals survived at 97.6% (41/42) and chief complaints at 95.2% (40/42), but medication facts fell to 47.6% (20/42; CI 33.4–62.3%). More than half were dropped or distorted.

4. Dangerous brand hallucinations:
   Under the Standardize prompt without a glossary, distortions more than doubled (4 to 9). The model didn’t recognize local brands and guessed similar-sounding Western drugs. In Scenario 1, Risek 40mg (omeprazole, for stomach ulcers) became risedronate 40mg, an osteoporosis drug that can itself cause severe stomach irritation and ulceration. In Scenario 2, Gabica 75mg (pregabalin) became gabapentin.

5. A simple in-context fix:
   With the 10-word glossary, brand hallucinations dropped to zero in this benchmark. The model correctly recorded Risek (omeprazole) 40mg and Gabica (pregabalin) 75mg.

---

Question 3: The Path Forward

The findings point to three targeted fixes, none of which need heavy retraining:

1. Formulary grounding for local brands:
   Add a lightweight lookup that matches the doctor’s words against a national formulary (such as Pakistan’s PharmaGuide) and puts the generic mapping into the prompt. In this benchmark, a small glossary eliminated brand hallucinations at no training cost, and a full formulary should extend that, though it needs testing at scale.

2. Role-gated, two-pass extraction for speaker errors:
   To keep patient self-treatments (7Up, Eno) out of the doctor’s plan, split generation in two. Pass 1 writes the Subjective section using only turns labeled [Patient]. Pass 2 writes the Plan using only turns labeled [Doctor]. This should sharply reduce cross-speaker errors, since the model can no longer see the wrong speaker’s words.

3. Targeted DPO for symptom idioms:
   To stop small models from dropping phrases like "seene vich khich paindi ae" or "talway jalte hain", align them with Direct Preference Optimization (DPO) on paired examples. The training should reward keeping the patient’s descriptive sensations over aggressive summarizing or jumping to Western diagnoses.

4. What not to do first: full end-to-end fine-tuning:
   The model’s core clinical reasoning looked sound, and retraining on scarce South Asian medical data risks catastrophic forgetting. In-context grounding and role-gated pipelines address these specific gaps more directly.

---

Limitations

• The dialogues were written to resemble typical outpatient visits, with AI assistance by a medical student, not recorded from live consultations.
• A single evaluator scored the notes, following a frozen rubric with automated consistency checks.
• Only a 4B-parameter model was tested, and larger models may behave differently.

---

Repository Artifacts

• README.md: Overview, full answers to required questions, key findings, and limitations.
• evaluationnotebookran.ipynb: Executed Google Colab notebook demonstrating model loading, prompting, and inference.
• consultation_dialogues.json: 7 full clinical consultation scenarios with tagged ground-truth clinical facts.
• labels.csv: Complete 210-row dataset of scored facts with labels and evaluator notes.
• raw_results.jsonl: All 42 generated clinical SOAP notes from the model.
• application_responses.txt: Clean, plain-text copy of the application responses.