Computational Clinical Reasoning
Every patient deserves deeper reasoning.
Clinical Superintelligence explores how AI can connect physiology, evidence, and the unfolding patient timeline into a more complete understanding of clinical care.
PropositionConnecting physiology, evidence, and time to make clinical reasoning deeper, clearer, and testable.
The next frontier is the depth of understanding.
Clinical information arrives in fragments: observations, monitor readings, laboratory results, treatment responses, and notes written at different times. Understanding a case requires connecting those fragments with physiology and relevant evidence.
AI creates an opportunity to explore more explanations, examine more interactions, and make assumptions easier to inspect.
Our ambition is clinical reasoning that is interdisciplinary, temporally precise, mechanistically grounded, and open to challenge.
- aInterdisciplinary
- Drawing on every specialty a case touches, rather than the one it first presented to.
- bTemporally precise
- Treating when something happened as evidence, alongside what happened.
- cMechanistically grounded
- Explaining a trajectory through physiology, not only through pattern and precedent.
- dOpen to challenge
- Showing every input, so that any conclusion can be questioned and corrected.
Six capabilities we intend to build and test.
These describe the research approach, not a deployed product. Each is a capability we aim to develop, evaluate against real clinical records, and expose to expert criticism.
- 01
Reconstruct the timeline
Connect observations, interventions, and responses. Preserve gaps in the record.
A missing interval is information. It should remain visible, not be smoothed over.
- 02
Reason through physiology
Explore the mechanisms that could explain a patient’s trajectory.
From ventilation to oxygen delivery to tissue demand: each step stated, each step checkable.
- 03
Connect disciplines
Bring relevant knowledge across specialties into a shared analysis.
Neonatology, cardiology, respiratory physiology and pharmacology rarely meet in one note.
- 04
Test competing explanations
Compare hypotheses and identify observations that would distinguish them.
The most useful output is often the next measurement that would settle a question.
- 05
Expose uncertainty
Separate measurements, reported observations, inferences, assumptions, and unknowns.
Five categories, carried through every figure and every claim.
- 06
Make reasoning auditable
Link conclusions to their evidence, model inputs, and limitations.
A conclusion without a trail back to its inputs cannot be reviewed, only believed.
Thirty-two minutes after birth
A neonatal oxygenation timeline shows both the promise of computational reasoning and the need to test its assumptions.
Illustrative retrospective case based on a supplied narrative and an AI-generated analysis. Original clinical records have not been independently reviewed for this website. Identifying details have been removed and times are given relative to birth.
Showing only what the supplied narrative reports. Nothing is drawn between readings.
- Reported reading
- Reported time window
- No record supplied
- Reported intervention
- Proposed reconstruction (illustrative)
- Assumption
What the narrative reports
- A dusky appearance is noticed.
- An SpO2 reading of around 50% is reported.
What the AI analysis inferred
- Arterial oxygen tension (PaO2, the partial pressure of oxygen in arterial blood) of 21.2 mmHg.
- A “systemic extractable margin” of 0.37-0.42 and an oxygen deficit of 5.0-5.4 mL O2/kg/min.
- A cerebral margin of 1.02, described as “knife-edge” but adequate.
Assumptions behind the numbers
- A specific oxygen-haemoglobin dissociation curve, used to convert saturation into PaO2.
- Whole-body oxygen consumption and cardiac output values not taken from this infant.
- Cerebral oxygen consumption of 1.8 mL O2/100 g/min, and cerebral blood flow rising from 40 to 80-85 mL/100 g/min.
- That a single reading at this level was accurate and representative.
Evidence that would help evaluate it
- Device, probe site, and signal quality for the reading.
- Heart rate at the time, which the supplied narrative does not report.
- Any blood gas or lactate measurement, and repeat readings.
Inferred values are unvalidated claims from the source analysis, shown so they can be examined. They are not patient findings or clinical guidance.
Claims under examination
Numbers as stated by the AI-generated source analysis, with what each one rests on. None has been validated; they are listed here to be examined, not used.
| Claim | Source value | Rests on | Status |
|---|---|---|---|
| Duration of hypoxic exposure | 22-24 min (minute 5 to ~28.5) | Assumed start at minute 5; assumed crossing time at 64% | InferredInferred from assumptions |
| SpO2 during minutes 5-17 | ~50% plateau | One later reading, extended backwards | Assumed |
| PaO2 estimates | 21.2, 24.4, 57.5 mmHg | Choice of dissociation curve; accuracy of reported readings | InferredModel-dependent |
| Systemic extractable margin | 0.37-0.42, 0.71, 2.56 | Proposed metric; assumed oxygen consumption and cardiac output | InferredMetric not validated |
| Systemic oxygen deficit | 5.0-5.4 and 2.47 mL O2/kg/min | Same inputs as the margin | InferredModel-dependent |
| Cumulative systemic oxygen debt | ~91 mL O2/kg | Deficits integrated over assumed plateaus | InferredInferred from assumptions |
| Cerebral oxygen consumption | 1.8 mL O2/100 g/min | Population value, not measured in this infant | Assumed |
| Cerebral blood flow | 40, rising to 80-85 mL/100 g/min | Assumed compensatory vasodilation | Assumed |
| Cerebral margin | 1.02 at 50%; 1.86 at 58% | Assumed flow and consumption, plus PaO2 estimates | InferredHighly sensitive to assumptions |
| Aerobic supply threshold | SpO2 64% | Output of the source model | InferredNot a validated threshold |
| FiO2 from 3 L/min nasal cannula | ~50% | Fixed flow-to-fraction conversion | Assumed |
| Neurological outcome | Injury “almost certainly” absent if heart rate ≥ 100 bpm | Heart rate not in the narrative; all inputs above | UnknownCannot be established from this source |
The breakthrough is making deeper reasoning possible, inspectable, and testable.
Retrospective modelling can reveal useful questions and competing explanations that a brief record does not make obvious. Here, it turned a handful of reported details into a detailed physiological account and showed precisely where that account depends on things nobody observed.
Its conclusions are only as strong as the quality of the record, the validity of the model, and the robustness of its assumptions. This page makes no judgement about the infant’s outcome or about the care provided.
Every conclusion should show what supports it.
A reconstruction can be detailed and still be wrong. The safeguard is structure: every statement carries its evidential status, and every number can be traced to the inputs that produced it.
Measured
Values captured by an identified instrument or test.
Example: A time-stamped oximeter export with device and probe site recorded.
Reported
Observations described in a narrative or record.
Example: “Dusky, SpO2 around 50%” written in an account of events.
Inferred
Interpretations drawn from available evidence.
Example: An arterial oxygen tension estimated from a saturation reading.
Assumed
Inputs introduced to make a model possible.
Example: A population value for cerebral oxygen consumption.
Unknown
Information the available evidence cannot establish.
Example: Saturation during an interval when nothing was recorded.
In the case study, every observation is classed as reported, because the primary records have not been independently reviewed.
A research workflow
The workflow we intend to develop and evaluate. It describes how a case should be examined; it has not been completed for the example on this site.
Assemble and reconcile the record
Gather notes, observations, monitor data, and results onto one time axis, reconciling clocks and duplicates.
Identify gaps and conflicting observations
Mark intervals with no data and readings that disagree, before any explanation is attempted.
Construct competing mechanistic explanations
Generate several physiological accounts of the trajectory, not one preferred story.
Link claims to primary evidence
Attach every claim to the observation, source, or assumption it depends on.
Test sensitivity to uncertain inputs
Vary assumptions across plausible ranges and see which conclusions survive.
Seek expert review and empirical validation
Put the analysis in front of clinicians and test models against outcomes they did not see.
Sensitivity analysis
Does a conclusion persist when plausible assumptions change?
Most clinical models need inputs nobody measured: an organ’s oxygen consumption, a blood flow, a curve relating one quantity to another. Sensitivity analysis asks what happens to a conclusion when each of those inputs moves across the range it could plausibly take.
A conclusion that holds across the whole range is robust. One that reverses within it is a hypothesis, and the analysis has identified exactly which measurement would settle it.
25 of 25 combinations support the conclusion
Robust across this range.
- Values chosen for the model
- Conclusion holds
- Conclusion reverses
Research agenda
Four proposed directions.
This is a research agenda: the questions we intend to pursue. It is not a record of completed work, and no results are claimed.
Temporal clinical reasoning
Understanding trajectories and intervention responses.
Open questions
- How should irregular, incomplete, time-stamped records be represented for reasoning?
- How can a system distinguish a response to treatment from coincident change?
Mechanistic model integration
Connecting language models with explicit physiological models.
Open questions
- When should a language model hand a question to an explicit, inspectable model?
- How are model limits and validity ranges carried into the final explanation?
Evidence provenance
Tracing claims to sources, observations, and assumptions.
Open questions
- What is the minimum provenance a clinical claim needs to be reviewable?
- How should evidential status propagate through a chain of inferences?
Evaluation and calibration
Testing accuracy, uncertainty, failure modes, and clinical usefulness.
Open questions
- Do stated uncertainties match how often conclusions turn out to be wrong?
- Does deeper reasoning change clinicians’ decisions, and for the better?
Toward a more complete understanding of every case.
This work needs people who rarely share a room: clinicians who know what a record leaves out, physiologists who can say when a model is wrong, machine learning researchers who understand failure modes, and scientific computing engineers who can make models honest and reproducible.
We welcome collaboration in principle across medicine, physiology, machine learning, and scientific computing. Formal channels will be announced here as the initiative develops.