Platform

From pathology result to clinical action, without the avoidable delay.

The failure mode isn't unstructured pathology. It's clinical latency — the gap between a clinically significant result and the action it should trigger. Echo Pathways closes that gap with four governed steps, and shows its working at each one.

01 · connect

Retrieve context

A read-only FHIR connection pulls the relevant pathology and patient context from the system that already holds it. We don't ask the hospital to move its data.

02 · interpret

Structure the evidence

Reads FHIR, HL7 v2, PDF, image and free text. Normalises units, scales and censoring conventions; recognises amended and superseded reports — always preserving the link to the source.

03 · apply

Run the protocol

A deterministic rule engine evaluates the patient's active, signed-off protocol: trend detection, threshold crossings, and results that were expected but never arrived.

04 · alert

Alert & assign

A prioritised recommendation is assigned to the responsible clinician, with the source, the rule and the reason — and stays visible until a human reviews it.

The oversight worklist

One screen: what needs a human, and who owns it.

The absence safety net

The result that never arrived is the one nobody is looking at.

Most surveillance tools react to data. The dangerous case is the opposite: a protocol expected a repeat marker in 12 weeks, and nothing came. There is no message to overlook, no row in an inbox, no alert to dismiss — only silence.

Echo Pathways tracks expected results as first-class objects and runs a clock against them. An overdue expected result is treated as high-severity, and escalates. This is the invariant we protect most carefully.

High Overdue expected result

No npm1_mrd received for 84 days; protocol expects a repeat every 12 weeks (overdue-expected-result).

rule · overdue-expected-result

reason · days_since_last > interval

assigned · Dr A. Goncalves

Evidence & provenance

Every recommendation can be replayed.

Open any action and you get the rule that fired, the exact inputs it was given, the source document it came from, and a hash of what we read. Re-fetch, re-hash, compare — the decision reconstructs. "Why did this fire?" has an answer, not an interpretation.

Action detail with the evidence sheet open, showing the rule, its inputs, the source document and provenance. — **Action & evidence** — the rule, its inputs, the source and the hash, on one sheet.

Patient timeline: NPM1 MRD on a log scale with treatment phases overlaid. — **Patient timeline** — the marker on a log scale, treatment phases overlaid, each point traceable to its source.

The role of AI

AI structures the pathology. Approved clinical rules determine whether action is required.

AI supports

Reading different report formats — FHIR, HL7, PDF, image, free text
Extracting the value, and identifying test, marker, specimen and assay
Normalising units and scales; flagging amended or superseded reports
Recognising uncertainty — low confidence goes to a human, never a guess

Approved clinical logic determines

Whether a protocol applies and a threshold is crossed
Whether a test is overdue or a trend needs review
Who is alerted, and by when
Signed off by clinical leaders before it runs

A pathology document is uploaded; the model reads a value of 1.16% at 99% confidence and routes it to a verification queue with the evidence snippet; a human confirms and commits it; only then does the deterministic rule fire and raise a relapse-trend alert. — The boundary, in one screen. The model **reads** 1.16% off an uploaded image and shows the snippet it read it from — then waits. A human confirms and commits. Only then does the deterministic rule evaluate the trend (prior 0.0007%, log-rise 3.219 against a 0.5 threshold) and raise the alert. The model never decided anything.

How protocols are governed What it does not do