BME363 ChatEHR ECA Sample: LLMs in Clinical Care Singapore

For information and reference only – not to be submitted as your own work.

This extract shows how a strong answer to the SUSS BME363 end-of-course assessment on ChatEHR could be structured. It covers parts 1a, 1b, 1c and 1e in full, with outline bullets for 1d and 1f. The original task is on the BME363 ChatEHR ECA question page, and our guide to answering the BME363 ChatEHR ECA explains how marks are usually earned.

Every ChatEHR figure below comes from the Stanford paper (Shah et al., 2026). Check each number against the paper before you reuse it, because the preprint has been revised.

Question 1a: Limitations of LLMs in medicine and requirements for safe use

Limitations

• Unsupported content. Shah et al. (2026) split unsupported claims in summaries into hallucinations (not in the record) and inaccuracies (contradicting the record). Even after deployment, summaries averaged 0.73 hallucinations and 1.60 inaccuracies per generation.
• Weak handling of time and attribution. The authors list the treatment of time and the attribution of facts as unresolved, which matters when a chart spans many years.
• Evaluation gaps. Benchmarks cannot anticipate every task clinicians will ask of a chat tool once it is live.
• Human factors. Automation bias and weaker verification habits are risks the paper names directly.
• Equity. Uneven data can create performance gaps even when no demographic inputs are given.

Requirements for safe and effective clinical use

• Task-specific evaluation against clinician-labelled data, repeated whenever the model changes.
• Ongoing monitoring of integrity, performance and impact after launch.
• Outputs that point back to the source record, so a clinician can check them quickly.
• Training for users, and governance that includes ethics review.
• Keeping a clinician accountable for the final decision.

In your own answer, link each limitation to its matching requirement, then judge which requirement is hardest to meet. That judgement is where the analysis marks sit.

Question 1b: Capabilities and evaluation framework of ChatEHR

Capabilities

ChatEHR lets LLMs work on a patient’s full, multi-year record inside the electronic health record. The paper describes three components: data orchestration (choosing data by time range, encounter type and location), context management, and LLM routing (selecting a model by token count and splitting very large contexts into chunks). It is model-agnostic, with 18 LLMs from several providers available through Stanford’s SecureGPT when the paper was written. It supports developer exploration, scheduled or triggered automations, and a chat interface embedded in the EHR.

Evaluation framework

• Benchmarks. MedHELM, with 121 clinician-validated tasks, informed which models to include.
• Automations. Each has a gold-standard dataset with clinician-adjudicated labels, and benchmarks are rerun when models change.
• User interface. Because tasks are not known in advance, the team combines quantitative measures, such as unsupported-claim rates, with interaction logs, task categorisation and user feedback.

A good answer compares the two routes: automations can be tested against a fixed truth set, while the open chat interface needs continuous monitoring. Say why in one sentence, then move on.

Question 1c: The concept of automations, with one example

An automation is a fixed combination of prompt and code that performs one defined task across many patient records. It can be triggered by an event, run in batch, or called on demand. Unlike free chat, its inputs and expected outputs are stable, so it can be validated and monitored like any other clinical tool.

Example: inpatient hospice eligibility review

This is one of the seven automations the paper reports in routine use. The tool reviews inpatient charts and flags patients who may qualify for hospice, so staff look at a short list rather than every chart. The paper reports feedback that was 81% positive and an estimated time saving of roughly 6,570 hours a year (about three full-time staff). It also notes that 20% of confirmed flagged patients were referred.

Explain why this suits automation: the question is narrow, the data elements are fixed, and a clinician can confirm each flag. Then name one risk, such as a missed eligible patient, and how monitoring would catch it.

A note on style: write in full sentences under each heading, define ChatEHR once at first use, and keep bullets for lists of evidence rather than for your argument. Markers reward a clear line of reasoning more than a long list.

Question 1d: Governance and team design (outline)

• Responsible AI Lifecycle established in 2023: risk tiering, assessments, ethics review and monitoring.
• Data science staff (engineers, product managers, data scientists) sit inside the IT organisation, with access to researchers.
• Staged build: sandbox, concept, prototypes, pilot, 150 clinical users, then system-wide rollout.
• Argue why embedding the team matters for speed and control (the “build-from-within” idea).

Question 1e: Value assessment of automations and the interface pilot

The table is a compact comparison. All values are estimates reported by the authors, not measured savings.

Factors influencing estimated value

• Fit with an action-oriented workflow, so results reach staff where they already work.
• Stable prompts and data, and a truth set to check against.
• Small time savings are “soft” because the minutes are often not reused.
• Finance, clinical and patient views of value differ, and patient benefit is hard to quantify.

Conclude with a judgement: the automations give firmer value, the interface gives wider reach but weaker evidence.

Question 1f: Challenges and limitations (outline)

Group the points into technical, human and organisational challenges so the answer reads as an analysis, not a list.

• Benchmarks do not cover every user task.
• Users need ongoing training in prompting and data selection.
• Expectations set by consumer chatbots (citations, fluid dialogue, scanned documents).
• Time handling and fact attribution still unresolved.
• Change management across services, and equity risks.
• Add your critical view: which limitation would stop you adopting it first?

If you need support with the full six-part answer, the SUSS assignment support page explains how we work, and our artificial intelligence assignment experts cover LLM and healthcare AI topics. You can also read student reviews before deciding.

About the expert

Gabriel Tan holds a PhD in Computer Science from the National University of Singapore. His expertise lies in the fields of artificial intelligence, machine learning, data analytics, and natural language processing.

References

Shah, N. H., Ambers, N., Pandya, A., Keyes, T., Banda, J. M., Nallan, S., Lugtu, C., Trotsyuk, A. A., Bedi, S., Unell, A., Fuentes, M., Grolleau, F., Jain, S. S., Chen, J., Dash, D., Char, D., Sharma, A., McElfresh, D., Scully, P., . . . Pfeffer, M. A. (2026). Adoption and use of LLMs at an academic medical center (arXiv:2602.00074). arXiv. https://doi.org/10.48550/arXiv.2602.00074