Skip to content

Research prototype · Clinical reasoning over time

TIMELY AGENT answers questions about a patient's record using only what had been recorded by the decision time. A deterministic engine applies the guideline's criteria, and every verdict quotes the passage it rests on.

Synthetic records only · Not a medical device

The demo at 02 Jan 12:00. The answer reads AKI detected, stage 1, anchored at 02 Jan 12:00, with its guideline sources; the chart shows creatinine rising from 1.0 to 1.6 mg/dL.The demo at 02 Jan 12:00. The answer reads AKI detected, stage 1, anchored at 02 Jan 12:00, with its guideline sources; the chart shows creatinine rising from 1.0 to 1.6 mg/dL.
Teaching patient A, a synthetic record, at three decision times. Screens from the demo.

01The problem

Hindsight is the easiest error to make with a patient record.

A result reported at noon cannot explain a decision made at ten, yet a record read after the fact mixes what was known at the time with what came later.

TIMELY AGENT fixes the decision time first and reads only what had been recorded by then. A later result stays out of view, and a criterion that cannot be judged from what is there is reported as not assessable rather than guessed.

02How it works

Five parts, each of which can be checked.

  1. 01

    Knowledge taken from the guideline

    Each condition is a knowledge pack: criteria, time windows and data-quality policies. Every rule cites a verbatim passage of the published guideline, with its page and a hash of the document it was checked against.

  2. 02

    The record as it stood

    The record is read at the decision time. A result recorded after it is out of view, however much it would have helped.

  3. 03

    A deterministic engine

    The engine evaluates every criterion as met, not met or not assessable, and names the data-quality policies it applied. The same record at the same decision time always gives the same verdict.

  4. 04

    A language model that never sees a value

    The model chooses which tools to call and writes its answer with placeholders. The values are filled in on the machine that holds the record, and every request that leaves it is fingerprinted and logged.

  5. 05

    A ledger for every conclusion

    The record, the knowledge, the engine and the result each carry a hash. A replay file reruns the engine on a synthetic record and checks every link.

03The demo

Everything above can be tried in the demo.

The demo opens on synthetic teaching records. It runs in the browser as a replay of the engine's stored results; the live agent runs on a local server. The images below are screens from it.

01

Every verdict shows its evidence

The Evidence tab lists the guideline passages behind the criteria that were met, word for word, with the page they come from and whether a person has checked them yet.

The Evidence tab for Teaching patient A at 02 Jan 12:00: KDIGO passages quoted word for word, each with its locator and a verification badge.The Evidence tab for Teaching patient A at 02 Jan 12:00: KDIGO passages quoted word for word, each with its locator and a verification badge.

02

Take one result away

Leave an observation out and the engine reads the record again without it. The reading takes the verdict's place, marked hypothetical, until you put the observation back.

The record at 02 Jan 12:00 with the 02 Jan creatinine left out: the engine's reading without it is not assessable, shown as a hypothetical.The record at 02 Jan 12:00 with the 02 Jan creatinine left out: the engine's reading without it is not assessable, shown as a hypothetical.

03

Try another value

Drag the latest result across its range. The marks show where the engine's verdict changes: had the 02 Jan creatinine been 2.0 mg/dL, the reading would have been stage 2.

The record at 02 Jan 12:00 with the 02 Jan creatinine tried at 2.0 mg/dL: the engine's hypothetical reading is AKI detected, stage 2.The record at 02 Jan 12:00 with the 02 Jan creatinine tried at 2.0 mg/dL: the engine's hypothetical reading is AKI detected, stage 2.

04

Check the working

The ledger lists the hash of every link behind a conclusion: the record, the knowledge, the engine and the result. A replay file lets anyone rerun the engine and compare.

An answer's ledger: the hashes of the record, the knowledge files, the engine and the result, and a link to download a replay file.An answer's ledger: the hashes of the record, the knowledge files, the engine and the result, and a link to download a replay file.

05

Compare two runs

Ask the same question of a record with one value changed and follow both runs step by step, with each difference marked as coming from the value or from the model.

The two-runs card: Teaching patient A at stage 1 and A′, with the 02 Jan creatinine changed from 1.6 to 2.1 mg/dL, at stage 2.The two-runs card: Teaching patient A at stage 1 and A′, with the 02 Jan creatinine changed from 1.6 to 2.1 mg/dL, at stage 2.

04Knowledge

Two conditions, each read from its guideline.

One record can be read for acute kidney injury and for hyperkalaemia. Each condition has its own knowledge pack, its own engine and its own verdict.

Acute kidney injury

Stage 1 · Stage 2 · Stage 3

Increase in SCr by ≥0.3 mg/dl (≥26.5 μmol/l) within 48 hours
Chapter 2.1 · Recommendation 2.1.1 · p. 19 · snapshot 20db2b44ea22
Reads
Serum creatinine · Urine output · Initiation of renal replacement therapy · Anuria
Sources
  1. KDIGO Clinical Practice Guideline for Acute Kidney Injury, 2012
  2. KDIGO 2026 Clinical Practice Guideline for AKI and AKD, public review draft
  3. NICE NG148, Acute kidney injury: prevention, detection and management
  4. NHS England algorithm for detecting AKI from serum creatinine changes, 2014
  5. UK Kidney Association Clinical Practice Guideline: Acute Kidney Injury, 2019
  6. ADQI 16 consensus report on acute kidney disease and renal recovery, 2017
Standards
LOINC 2160-0 · UCUM · OMOP CDM v5.4

Acute hyperkalaemia

mild · moderate · severe

It is further classified by severity into mild (5.5-5.9 mmol/l), moderate (6.0-6.4 mmol/l) or severe (≥ 6.5 mmol/l).
Introduction · p. 8 · snapshot c8eef6a55a95
Reads
Serum potassium · ECG features of hyperkalaemia
Sources
  1. UKKA Clinical Practice Guideline: Management of Hyperkalaemia in Adults, 2023
Standards
LOINC 2823-3 · UCUM

Passages are checked word for word against a hashed snapshot of the published guideline. Both knowledge packs are research drafts awaiting clinical sign-off.

The demo's Knowledge view: acute kidney injury v0.2 and acute hyperkalaemia v0.1, both drafts not clinically validated, with their sources, passages, criteria and states, and the measured concepts each one reads.The demo's Knowledge view: acute kidney injury v0.2 and acute hyperkalaemia v0.1, both drafts not clinically validated, with their sources, passages, criteria and states, and the measured concepts each one reads.
The Knowledge view in the demo: what each condition is built from, and the measured concepts each one reads.

05Boundaries

What it does not do.

  • It does not diagnose or prescribe.

    It reports what a guideline's criteria say about a record at a decision time. Decisions stay with clinicians.

  • It does not guess.

    When the record cannot answer, the verdict is not assessable, and the answer says what is missing.

  • It does not show the model a value.

    The language model works with placeholders. Values are filled in on the machine that holds the record.

  • It is not a medical device.

    TIMELY AGENT is a research prototype, and every public example uses synthetic records.

06Research

The same engine can grade a language model.

TIMELY began as a framework for benchmarks of temporal clinical reasoning. Its arena poses questions to a language model at decision times on synthetic cases and scores each answer against the engine.

Arena in development · synthetic cases only

Status
Does the model reach the engine's verdict at this decision time?
Stage and anchor
The grade, and when the episode began.
Evidence
Rerun on only what it cites, does the engine agree with it?
Calibration
How well its stated confidence matches how often it is right.
Asking
Whether it asks for earlier history when the answer depends on it.

See it read a record.

Open a synthetic teaching record, set the decision time and ask it a question. The demo runs in your browser.

Try the demo