Jonahnki/Clinical_reasoning_under_resource_constraint

Resource-Aware Clinical Reasoning: A Paired-Vignette Evaluation of MedGemma 1.5 4B

1. The blind spot

Medical benchmarks ask what a clinician knows. In low- and middle-income settings, the harder question is what a clinician does when the textbook investigation is not in the building, or is in the building and beyond what the family can pay.

In my own experience, the disparity between theoretical medicine and clinical reality is wide, and it widens in LMIC conditions. I saw one case worked around a referral delay. The patient passed through several hospital facilities, and each reached a different diagnosis, depending partly on the facility's resources and partly on the efficacy of the investigation it had to use. Relatives went from street to street and to extended family for money while the misdiagnoses accumulated. We lost the patient.

Medical language models learn from literature written for well-resourced settings. AfriMed-QA, the Pan-African benchmark included in the MedGemma 1.5 evaluation, tests medical knowledge recall. It does not test whether a model changes its plan when the default investigation is out of reach. The gap between knowing the right test and choosing the right plan for the facility in front of the patient is what this dataset measures.

Methodology

Vignettes. Twenty clinical cases, each written as a pair. Variant A gives the standard presentation. Variant B repeats the case and adds one sentence stating a resource constraint, such as no on-site ultrasound at a rural secondary hospital. Each vignette carries an answer key with three fields: the well-resourced default, the resource-aware action, and any locally harmful practice. Vitals and laboratory values in the stems are synthetic.

Models.

• google/medgemma-1.5-4b-it, a medically trained model
• Qwen/Qwen2.5-3B-Instruct, a generalist baseline of similar size

Both models ran single-turn with greedy decoding and the same structured prompt: leading diagnosis, immediate next steps, prioritised investigations. Precision: [state the precision the notebook used, for example bf16]. Eighty generations took about 14 minutes on a Colab T4 GPU.

Scoring. One rater, the vignette author, scored the responses by hand against the vignette key, using three binary columns:

Column	Intended meaning of a 1
axis1_recommends_unavailable_investigation	The response recommends an investigation or treatment that is unavailable in the setting
axis2_names_bedside_cue	The response names a clinical sign or bedside step from the resource-aware key
axis3_mentions_cost_or_availability_unprompted	The response raises cost, distance, or availability without being asked

Availability was judged against the rater's knowledge of practice in Nigeria, not only against the constraint sentence in variant B. Variant A responses were therefore scored against local reality too, which is why variant A rates on axis 1 are high. The rater also wrote a free-text note on 55 of the 80 rows.

Results

Counts as recorded in scoring_template.csv, with percentages in parentheses.

Model	Variant	Axis 1	Axis 2	Axis 3
MedGemma 1.5 4B	A	18/20 (90%)	18/20 (90%)	19/20 (95%)
MedGemma 1.5 4B	B	15/19 (79%)	17/19 (89%)	17/19 (89%)
Qwen2.5 3B	A	9/19 (47%)	9/19 (47%)	10/19 (53%)
Qwen2.5 3B	B	12/20 (60%)	12/20 (60%)	12/20 (60%)

Unscored responses. Two responses carry no score: vignette 3, MedGemma, variant B, and vignette 5, Qwen, variant A. The MedGemma response ends mid-sentence, inside a long list of prioritised investigations. The Qwen response is a single line naming severe infection and a blood culture. Those two groups have n = 19. Scoring the missing responses as 0 or as 1 moves the axis 1 rate for MedGemma B between 75% and 80%, and for Qwen A between 45% and 50%. No comparison in this card changes.

Vignette 12. The case has no distinct constraint variant, so its two rows carry the same prompt. Removing it leaves the pattern unchanged: on axis 1, MedGemma scores 17/19 in A and 14/18 in B, and Qwen scores 9/18 in A and 12/19 in B.

What the paired data show. Within each vignette, the constraint sentence changed MedGemma's axis 1 score in 3 of 19 pairs, all three falling from 1 to 0. Qwen's axis 1 score changed in 8 of 19 pairs, rising in 5 and falling in 3. A movement of 3 to 8 vignettes out of 19 sits inside the noise of a single-run, single-rater study. The data show no consistent adaptation to the stated constraint by either model.

Response length. MedGemma's responses average 894 characters and Qwen's average 427. A longer plan has more places to include an investigation the facility cannot run.

What the rater's notes show. The clearest evidence sits in the free-text notes, and it matches the blind spot described in section 1.

• MedGemma variant B responses still included investigations the rater judged unavailable or unnecessary in local practice: point-of-care ultrasound (vignette 4), rapid antigen tests (vignette 5), a TORCH screen (vignette 10), an unnecessary investigation list (vignette 11), and troponin, procalcitonin, and D-dimer (vignette 18).
• For vignette 5 in variant A, the rater noted that a long investigation list would push a patient to abscond from care.
• Some MedGemma plans were well matched to the setting. The rater rated the vignette 7 plan as readily available out of pocket in both variants.
• Qwen recommended fewer investigations but included blood gas analysis (vignettes 4 and 18), endoscopy (vignette 20), MRI or CT (vignettes 6 and 8), and spirometry (vignette 20). It also reached wrong or unrealistic diagnoses, such as gastroesophageal reflux disease for a classic pyloric stenosis presentation in vignette 1. A model that recommends less scores better on axis 1 without reasoning better.

MedGemma 1.5 4B scores 16.4% on MedXpertQA (text only), so some of its errors may come from the limits of small-model reasoning and not from resource blindness. The design cannot separate the two.

Scoring audit

Review of the scored file after the run found two problems, and the results above should be read with them in mind.

1. The three columns move together. Of 78 scored rows, 68 carry identical values on all three columns: 50 rows are all 1 and 18 rows are all 0. Three separable behaviours would not agree that often. The recorded values behave like a single judgment of overall response quality.
2. Axis 1 direction was not applied consistently. Axis 1 is defined as a failure indicator, where 1 means a recommendation the setting cannot support. On several rows the note and the score disagree. For example, the notes for vignette 4 (Qwen, A), vignette 5 (MedGemma, B), vignette 6 (Qwen, B), and vignette 14 (Qwen, A) describe investigations unavailable or unfeasible locally, and each row scores 0. The note for vignette 12 (MedGemma, B) says all recommendations are locally available, and the row scores 1.

The percentages in the results table are therefore ratings as recorded, not validated measures of three separate behaviours. The paired counts, response lengths, and rater notes carry the findings. A second pass by the same rater, or a second rater, with a one-line decision rule for each column, would make the axis rates comparable.

3. Path forward

The results point to a data problem before an architecture problem. A model that names the right bedside sign, acknowledges the constraint in a sentence, and then lists the unavailable test anyway has learned both halves of the answer and has not learned to choose between them. Three changes address that choice directly.

1. Curate facility-tiered cases. Training and evaluation sets need the same presentation written for several facility levels, from a primary health centre to a tertiary teaching hospital, each with its own correct action. Clinicians who practise at those levels should write and review them. The WHO Model List of Essential In Vitro Diagnostics gives a defensible starting point for which tests belong at which level.
2. Train on the trade-off, not the fact. Preference pairs in which the resource-aware plan beats the textbook plan under a stated constraint teach the behaviour the benchmark measures. Supervised fine-tuning on medical knowledge alone would not, because the knowledge is already present.
3. Ground plans in local guidance and cost. Retrieval over Nigeria's national standard treatment guidelines and protocols built around bedside signs, such as WHO's Integrated Management of Childhood Illness, gives the model the facility-level defaults general medical corpora omit. The plan should also sequence investigations by what changes management and what the family can afford, since a long list is itself a cause of patients leaving care.

The next version of the benchmark needs facility-tier labels on every vignette, two or more raters from different practice settings, and one decision rule per scoring column written before any response is scored.

Limitations

• Twenty vignettes, one run per prompt, no resampling.
• The vignette author is also the sole rater, and scoring was not blinded to model.
• The three scoring columns are highly correlated and axis 1 direction was inconsistent on several rows; see the scoring audit.
• Availability was judged against the rater's knowledge of practice in Nigeria and does not generalise to every LMIC setting.
• Vignette 12 has no true constraint variant.
• MedGemma was tested text-only despite being multimodal.
• The MedGemma model card states that the model is more prompt-sensitive than Gemma 3 and has not been evaluated for multi-turn use. One prompt template was tested.
• Vitals and laboratory values in the stems are synthetic.
• The answer key was checked against public guidance where noted in the verification_note column and otherwise reflects standard teaching. It has not had independent clinical review.

Related work

MedGemma 1.5 4B scored 56% on AfriMed-QA against 48% for Gemma 3 4B, on a 25-question test set reported in the MedGemma 1.5 model card. The benchmark measures medical knowledge recall. This evaluation measures plan selection under a stated constraint.

Files

File	Contents
vignette_template.csv	The 20 vignettes, both variants, with answer keys
results.csv	Raw model responses
scoring_template_scored.csv	Hand-scored responses with rater notes
medgemma_resource_aware_eval.ipynb	Generation notebook

Reproduce

Run the notebook on a Colab T4 GPU with transformers>=4.50.0. Accept the MedGemma license on your own Hugging Face account before loading the model. Model weights are not redistributed here. Eighty generations take about 14 minutes.

Disclaimer

Research use only. Nothing in this dataset is clinical advice, and MedGemma outputs are not intended to inform patient care.