CardioAssist — A Clinician-in-the-Loop Decision Support System for Cardiac Differential

In the emergency department, cardiac presentations must be assessed fast, on dense, multi-source data. A clinician evaluating chest pain or breathlessness has to synthesise an ECG, an echo report, a troponin trend and the patient’s history within minutes – often at the end of a long shift and while managing several patients.

This is not a knowledge gap; experienced clinicians know the medicine. It is a context-and-workload gap that drives three recurring risks:

• Anchoring and premature closure: where the first plausible diagnosis crowds out competing life-threats

• Data overload: where a critical finding is buried and under-weighted

• Inconsistent documentation of reasoning: which weakens handover and audit

A digital solution was required because the problem is one of cognitive bandwidth under time pressure – where a fast, structured, always-available second read adds value that staffing or training alone cannot.

The clinicians most affected are emergency and duty physicians, who carry the diagnostic load; patients are the ultimate stakeholders, since a missed time-critical diagnosis (STEMI, pulmonary embolism, aortic dissection) causes direct harm.

NABH quality aspects impacted include:

• Assessment of Patients (AOP) – timely, systematic diagnostic assessment

• Care of Patients (COP) – safe emergency management

• Patient Safety & Quality Improvement (PSQ) – reducing diagnostic error and improving documentation

Relevant indicators:

• Time to diagnosis

• Diagnostic concordance

• Missed-critical-diagnosis rate

• Completeness of clinical documentation

CardioAssist is an in-house, clinician-developed clinical decision support tool for cardiac differential diagnosis – conceived, built and tested by a practising physician-executive (not outsourced).

The clinician uploads ECG images, echo reports and lab panels with the patient’s story. A vision-capable large language model extracts the findings into structured, editable fields (rhythm, intervals, ST/T, LVEF, troponin trend) and generates a ranked, reasoning-backed differential – life-threats first – each with supporting-versus-opposing evidence, recommended workup, critical alerts and an overall urgency rating.

Key Features:

• Multimodal ingestion

• Structured extraction that the clinician can review and correct

• Ranked differentials with transparent reasoning

• Recommended next steps and red-flag alerts

• Mandatory human-in-the-loop (it drafts, the clinician decides)

• Privacy by design – data is processed in memory per request, with no PHI stored

Every output is clinician-verifiable and editable.

CardioAssist was developed in-house as a clinician-led initiative, with Dr. Sameer Sital Raj as both clinical lead and developer – combining domain expertise with formal machine-learning training.

Rollout to date has been a controlled early evaluation rather than a hospital-wide deployment: the tool was run on 13 real emergency presentations across three dates (29 April, 11 July and 13 July 2026), with the treating emergency consultant reviewing each output.

[Insert your actual timeline – e.g., concept to working prototype in X weeks/months.]

Departments Covered:
[Emergency Department – extend to Cardiology / IPD as applicable]

Staff Involved:
[number of clinicians who used or reviewed the tool]

Internal Champions:
[name the consultant(s) who led the pilot and any quality/IT colleagues who supported it]

Because of the human-in-the-loop, no-PHI-storage design, onboarding is light – a clinician can use the tool with minimal training, since every value is editable and verifiable.

The plan is to conduct a proper planned pilot project for a couple of months, keeping DPDP and CDSCO compliances in mind.

CardioAssist’s impact to date is demonstrated through an early clinical evaluation rather than hospital-wide metrics.

Across 13 real emergency presentations over three dates, the treating clinician’s working diagnosis appeared in the tool’s top two ranked differentials in every case (top-2 concordance 13/13; 95% CI ≈ 77–100%), and the emergency consultant judged each ranked differential corroborative of their own assessment.

Operationally, the tool compresses the synthesis of ECG, echo and lab data into a structured, ranked draft in seconds, and produces documented reasoning (supporting-versus-opposing evidence, recommended workup) that improves the completeness and auditability of the diagnostic record.

On quality and safety, its life-threats-first design consistently surfaced time-critical diagnoses – pulmonary embolism, aortic dissection, cardiogenic shock – in the critical band even at low likelihood, directly targeting diagnostic-error and missed-critical-diagnosis risk.

This supports NABH digital-health accreditation aims by evidencing responsible AI use: mandatory human-in-the-loop, transparent auditable reasoning, no PHI storage, and explicit scope boundaries – the governance an assessor looks for.

It aligns with AOP, COP and PSQ standards.

Enablers:

The decisive enabler was clinician-led development: the tool was conceived and built by a practising physician-executive with formal machine-learning training, so clinical relevance and safe scoping were designed in, not retrofitted.

Vision-capable large language models made multimodal extraction (ECG, echo, labs) feasible without bespoke infrastructure.

A deliberate human-in-the-loop, no-PHI-storage architecture kept the governance and data-protection burden low, which in turn made real emergency-department testing possible quickly and safely.

Innovations:

• Ranked differentials with life-threats surfaced first

• Transparent supporting-versus-opposing evidence for every diagnosis so the reasoning can be audited rather than trusted blindly

• Fully editable outputs that make the clinician the decision-maker at every step

Challenges:

Auto-generated ICD-10 codes proved unreliable across cases – an honest limitation that reinforces the verify-every-value principle and drives a roadmap fix (a validated coding layer).

Other constraints:

• Dependence on input quality

• The need for a larger outcome-linked validation study

• Change-management to build appropriate clinician trust without automation bias

Top Learnings:

1. Scope discipline is a safety feature

Defining clearly what the tool is not (not a device, not a decision-maker, not a PHI store) is what makes it deployable and governable.

2. Transparency beats accuracy claims

Showing reasoning and opposing evidence builds justified clinician trust and enables audit.

3. Honest handling of limitations

Such as the ICD-coding errors we surfaced, strengthens rather than weakens the case.

For Other Hospitals:

• Keep a clinician in the loop and accountable

• Prefer tools that show their reasoning

• Minimise data retention

• Validate against your own case mix before scaling

• Treat AI output as a draft to verify, never an instruction