Notice. Illustration and research prototype. Not clinically validated, not a medical device, not for patient care. Example data are fictional.
Working register · v1 · August 2026

Ideas for AI in critical care

A scan of the possibility space, not a set of recommendations. Each entry is an idea that has been proposed, piloted, or deployed somewhere — described plainly. The positions are yours to form. This exists so they are formed before the conversation, not during it.

Recognising deterioration

01–05

The most crowded area, and the one with the worst record of external validation. Nearly every entry here works in the unit it was built in.

01

Continuous sepsis monitoring

SOFA, MEWS, qSOFA and twelve derived ratios recomputed on every change from values already recorded. Published thresholds applied to measured numbers — it reports what has left its band and in which direction. A second layer learns from the hours before confirmed cases and flags patients on a similar trajectory, with the alert burden set by what the unit can answer.

For reviewBuilt · v0.2 Every threshold carries its source and the populations it is unreliable in. The learned layer reports utility alongside AUROC — on generated data AUROC 0.93 but utility −0.35 at 36 alerts per case, reproducing the published external finding across 91 studies.
Technical detail
Bedside calculator with bidirectional bands, deterministic scores, ISCCM-grounded escalation tiers, 34 clinical regression tests. Bedside monitor(local, needs console.py running) Run report (generated at run time) ISCCM extract Evidence base
02

Acute kidney injury — staging and nephrotoxin review

KDIGO stage recomputed as values change, taking the worse of the creatinine and urine criteria rather than whichever was checked first. Alongside it a nephrotoxin count with the substitution for each — because the actionable part needs no forecast: which drugs this patient is on, whether the combination is avoidable, whether the dose still fits.

For reviewBuilt · v0.1 Deterministic staging with both criteria shown separately, drug-name resolution (gentamicin → aminoglycoside), and named substitutions. The vancomycin + piperacillin-tazobactam pairing carries OR 1.4 for AKI against vancomycin + cefepime — an avoidable combination, not just a warning.
Technical detail
Missing baselines are declared rather than assumed, since a patient admitted already injured looks normal against their own admission value. 33 tests passing. AKI review(local, needs console.py running) Staging & nephrotoxin table
03

Trajectory — direction and rate, not the current value

A monitor shows twelve current numbers and no direction. A shock index that has climbed 0.58 → 0.69 over four hours reads as “0.69” and looks normal — the trend is the signal and it is invisible. Computes slope, direction, time to the band edge at the current rate, and whether movement is accelerating, then orders the unit by who needs looking at.

For reviewBuilt · v0.1 Least squares over a six-hour window — a formula, not a model. Output is attention direction, not a forecast: it says which patient has changed and what moved, never a probability. Now carries a pharmacology layer: movement inside a drug's onset-to-peak window is labelled as treatment rather than flagged (urine rising 20 min after furosemide is the drug); a drug that should have worked and has not is itself a finding (no MAP response 25 min after increasing noradrenaline); and half-life is recomputed at the patient's clearance — vancomycin 6h normally, 29h at CrCl 12, clearing in six days. It also catches the commonest misattribution in critical care: a rising lactate on adrenaline is beta-2 stimulation, not necessarily worsening shock.
Technical detail
Ordering the unit is the product. 52 tests passing, including a guard that no numpy boolean escapes the engine.Trajectory view(local, needs console.py running) Drug table
04

Deterioration detection from camera alone

Vision on respiratory effort, posture, and colour, without a single lead attached. Attractive precisely where monitoring hardware is absent.

Awaiting data Camera-based inference, held for the same reason as the imaging block: the method is understood, the data is not in hand. Accuracy under real ward lighting is the unresolved part and can only be settled on footage from the units it would run in.
05

Post-ICU-discharge deterioration and bounce-back

Predicts who will return within 48 hours. Affects discharge timing, which affects bed availability — the binding constraint in most Indian units.

The ventilator

06–15

The best data in the hospital: dense, continuous, machine-readable waveforms that no human watches breath by breath. The strongest technical fit in critical care.

06

Patient–ventilator asynchrony detection

Catches ineffective efforts and double-triggering from waveform morphology. Under-detected by humans for the simple reason that nobody can watch every breath — a ventilated patient takes about 20,000 a day.

For reviewBuilt · v0.1 Geometric rules, not a trained model: each detection states its shape in one sentence and a panel can change the threshold or switch the rule off. Calibrated against normal breathing first — index 0.0 at every rate from 12 to 24/min, because a detector that fires on normal breathing gets switched off in a day and then misses what it was installed for.
Technical detail
Adapter layer with the six major makers declared and MEDIBUS serial settings recorded; stubs raise rather than fabricate, and no adapter exposes a write path. 55 tests passing.Command centre(local, needs console.py running) Detection rules Adapters
07

Weaning readiness screen

Ten published criteria run against recorded values on every ventilated patient, every day. The criteria are neither difficult nor disputed — the failure is that a busy morning means nobody screens, and each unnecessary ventilator day carries pneumonia risk, delirium, and cost.

For reviewBuilt · v0.1 Three commitments. Missing is not passing — a patient with no recorded pH is unassessable, not compliant. Judgement is asked, not computed — whether the cause is improving and whether secretions are manageable cannot be measured, so the screen stays incomplete until a clinician answers.
Technical detail
RSBI is shown, not enforced — AARC 2024 states it is not needed to determine readiness, so a patient with RSBI 120 still reads ready, and the departure from the classic teaching is stated on the output rather than buried. 48 tests passing.Command centre(local, needs console.py running) Criteria
08

Extubation failure risk

The inverse question, and the more consequential one. Readiness asks whether the patient can breathe without the ventilator; this asks whether they can protect and clear their own airway once the tube is out. Different predictors, which is why passing a breathing trial does not settle it.

For reviewBuilt · v0.1 A count, not a model: the three principal factors — weak cough, excess secretions, abnormal neurology — combine synergistically, and the published figures carry the discrimination on their own (3% reintubated with none, 100% with all three). Unmeasured is not absent: a patient whose cough was never assessed reads “Low risk — 2 factors not assessed”, never a clean low-risk verdict. Cuff leak handled separately, and an untested high-risk patient is itself a finding.
Technical detail
A cohort rate is labelled as such, not as a probability for the patient. 37 tests passing.Command centre(local, needs console.py running) Risk factors
09

Automated closed-loop ventilation

The machine adjusts its own settings continuously. Commercial systems exist; adoption is limited less by evidence than by willingness to hand over control.

Assessed · held Closed-loop control crosses the read-only boundary set for this build. Held rather than declined — it is a scope decision, not a judgement on the idea.
10

Lung-protective ventilation adherence

Computes predicted body weight from height and sex, derives the 6 mL/kg tidal volume target, and flags drift. Lung volume scales with height, not body mass — so setting by actual weight over-ventilates short and obese patients, every breath, for days.

For reviewBuilt · v0.1 Catches the specific error: a 152 cm woman at 85 kg has a target of 271 mL; set by actual weight she receives 510 mL, 11.2 mL/kg predicted. Also plateau pressure, driving pressure and PEEP.
Technical detail
A patient with no height recorded is counted as not adherent — without a target nothing can be within it. 26 tests passing.Ventilation adherence(local, needs console.py running) Targets
11

ARDS phenotype identification

Separates hyper- and hypo-inflammatory subtypes that respond differently to the same treatment. Scientifically strong, not yet actionable at the bedside.

Assessed · held Needs bedside biomarkers that do not exist outside research settings. Scientifically strong, not yet actionable.
12

Individualised PEEP titration

Optimal PEEP per patient rather than per table. Mechanistically compelling; no prospective trial has shown outcome benefit.

Assessed · held Mechanistically compelling with no prospective outcome benefit shown. Worth building as a comparison tool rather than a recommendation.
13

Self-inflicted lung injury risk in spontaneous breathing

Detects when a patient's own respiratory drive is causing harm. A newer concept; measurement remains the hard part.

Assessed · held Measurement is the unsolved part — oesophageal pressure or an equivalent is needed before the concept is actionable.
14

Ventilator-associated pneumonia early signal

Combines waveform, secretion, and lab trends. VAP is a headline quality metric in Indian ICUs, which gives the idea institutional pull.

Assessed · held Needs the unit’s VAP surveillance definition and secretion charting before the signal has anything to combine. The waveform half already exists.
15

Ventilator setup guidance for the untrained operator

Initial settings and escalation logic for a physician who has never been trained on the machine. Squarely aimed at the district-hospital gap.

Assessed · held Needs the ventilator models actually in use in the district units, and approved initial settings for each. The escalation logic is straightforward once those are named.

Imaging and the bedside signal

16–27

The most regulatorily mature category. Cleared products exist, including Indian ones, and the read is verifiable against ground truth — which is why it advanced faster than the rest.

16

Chest radiograph triage and abnormality flagging

Prioritises the film that needs reading now. Mature, cleared, and already commercial in India.

Awaiting data The method is settled and the input does not exist yet. This is a convolutional or vision-transformer task on real studies — not a rule engine — so it is held until there is a labelled corpus from the units it would run in, to train and validate against. Indian data specifically: a model trained elsewhere and dropped in is the failure this register documents repeatedly.
17

Endotracheal tube and line position verification

Confirms placement on the portable film automatically. Narrow, checkable, genuinely useful — the profile of a good first deployment.

Awaiting data The method is settled and the input does not exist yet. This is a convolutional or vision-transformer task on real studies — not a rule engine — so it is held until there is a labelled corpus from the units it would run in, to train and validate against. Indian data specifically: a model trained elsewhere and dropped in is the failure this register documents repeatedly.
18

Head CT haemorrhage detection and triage

Pushes the bleed to the top of the queue. One of the clearest cases of AI shortening time to intervention.

Awaiting data The method is settled and the input does not exist yet. This is a convolutional or vision-transformer task on real studies — not a rule engine — so it is held until there is a labelled corpus from the units it would run in, to train and validate against. Indian data specifically: a model trained elsewhere and dropped in is the failure this register documents repeatedly.
19

Pneumothorax detection on portable films

Time-critical and easily missed on a bedside AP film taken at 3 AM.

Awaiting data The method is settled and the input does not exist yet. This is a convolutional or vision-transformer task on real studies — not a rule engine — so it is held until there is a labelled corpus from the units it would run in, to train and validate against. Indian data specifically: a model trained elsewhere and dropped in is the failure this register documents repeatedly.
20

Lung ultrasound interpretation assistance

B-lines, consolidation, effusion — read by the machine. Ultrasound is cheap and portable, which makes this unusually well suited to resource-limited units.

Awaiting data The method is settled and the input does not exist yet. This is a convolutional or vision-transformer task on real studies — not a rule engine — so it is held until there is a labelled corpus from the units it would run in, to train and validate against. Indian data specifically: a model trained elsewhere and dropped in is the failure this register documents repeatedly.
21

Automated echocardiography for the non-cardiologist

Ejection fraction and filling assessment from a novice-acquired view. Converts a specialist skill into a bedside measurement.

Awaiting data The method is settled and the input does not exist yet. This is a convolutional or vision-transformer task on real studies — not a rule engine — so it is held until there is a labelled corpus from the units it would run in, to train and validate against. Indian data specifically: a model trained elsewhere and dropped in is the failure this register documents repeatedly.
22

Guided ultrasound acquisition

The system tells the operator how to move the probe. Solves acquisition rather than interpretation — usually the real bottleneck outside teaching hospitals.

Awaiting data The method is settled and the input does not exist yet. This is a convolutional or vision-transformer task on real studies — not a rule engine — so it is held until there is a labelled corpus from the units it would run in, to train and validate against. Indian data specifically: a model trained elsewhere and dropped in is the failure this register documents repeatedly.
23

Continuous ECG interpretation beyond arrhythmia

Ischaemia, electrolyte derangement, and structural disease inferred from the ECG the patient is already wearing.

Awaiting data Continuous-signal interpretation requiring a trained model over raw traces. Held until there is a recorded corpus with expert labels. The engineering is well understood; the labelled data is the constraint.
24

Continuous EEG interpretation for non-convulsive seizure

Non-convulsive status is missed constantly because reading continuous EEG requires a neurologist who is not there.

Awaiting data Continuous-signal interpretation requiring a trained model over raw traces. Held until there is a recorded corpus with expert labels. The engineering is well understood; the labelled data is the constraint.
25

Pupillometry and automated neurological assessment

Objective pupil response replacing a torch and a subjective note.

Awaiting data The method is settled and the input does not exist yet. This is a convolutional or vision-transformer task on real studies — not a rule engine — so it is held until there is a labelled corpus from the units it would run in, to train and validate against. Indian data specifically: a model trained elsewhere and dropped in is the failure this register documents repeatedly.
26

Non-contact vital sign estimation from video

Respiratory rate and pulse from camera alone. Interesting where monitoring hardware is scarce; accuracy under real ward lighting is the unresolved part.

Awaiting data Camera-based inference, held for the same reason as the imaging block: the method is understood, the data is not in hand. Accuracy under real ward lighting is the unresolved part and can only be settled on footage from the units it would run in.
27

Waveform quality and artefact rejection

Distinguishes a real arrhythmia from a lead that has come loose. Mundane, and upstream of every other idea in this register.

Treatment decisions

28–38

Where the field most wants to be and has least to show. The recurring failure: models trained on retrospective data learn to imitate the clinicians who generated it, including their errors.

28

Reinforcement-learning vasopressor and fluid titration

The famous version of this idea. Striking retrospective results; nothing has survived prospective testing.

Assessed · held Retrospective results are striking and nothing has survived prospective testing. Held as the cautionary case rather than as work.
29

Fluid responsiveness prediction

Whether this patient will respond to a bolus, without a passive leg raise. A daily question, asked dozens of times per unit.

30

Antimicrobial selection from local resistance patterns

Empiric therapy tuned to this unit's antibiogram and this patient's history. In India, where resistance is severe and local, the case is unusually strong.

31

Antimicrobial de-escalation

Flags when broad-spectrum cover should have been narrowed. Nobody disagrees that day-3 review should happen — the patient started on meropenem at 2 AM on Monday is still on it on Thursday because nobody revisited it. The failure is temporal, not clinical, which is exactly the kind a clock-driven engine fixes.

For reviewBuilt · v0.1 Day-3 review, culture-guided narrowing, duration against indication, and redundant cover. The asymmetry it enforces rather than notes: a biomarker may support stopping and must never justify continuing — a raised procalcitonin does not suppress the duration finding, because a marker that only ever extends therapy is worse than no marker. Procalcitonin is off by default on ISCCM feasibility grounds (6.7% against CRP at 93.3%).
Technical detail
Typical durations are shorter than most units prescribe, and a stated reason for continuing is treated as a decision rather than a lapse. 48 tests passing.Command centre(local, needs console.py running) Spectrum & durations
32

Sedation depth and daily interruption

Over-sedation is endemic and invisible without scoring. It drives delirium, ventilator days, and cost. The ABC trial paired awakening with breathing trials and cut ventilator days by 3.1 and one-year mortality by 14% — a larger effect than most drugs in critical care, from a protocol.

For reviewBuilt · v0.1 Four questions asked continuously: was a target set, when was the patient last scored, are they at that target, and is the drug the one the guideline recommends. The distinction carrying most of the value is deep by decision versus deep by drift — on a chart they look identical. Two demo beds sit at RASS −4: one reads critical (no indication, on a benzodiazepine), the other reads at target (indication recorded).
Technical detail
Depth below −3 without a recorded indication is the only critical finding in the engine. 45 tests passing.Command centre(local, needs console.py running) Targets & drug choice
33

Transfusion thresholds

Restrictive strategies are well evidenced and inconsistently followed. The gap is adherence, not knowledge — and the reason is that the threshold is not one number. A haemoglobin of 7.4 is above threshold in a stable septic patient and below it after cardiac surgery. On a screen they look identical.

For reviewBuilt · v0.1 Context-varying thresholds from TRICC, TRISS and TRICS-III, with active bleeding overriding the number entirely — a haemoglobin measured during bleeding describes the past. Flags both directions: restrictive is not withholding, so below-threshold and untransfused is a finding too, unless a reason is recorded. Enforces single-unit practice, the half of restrictive transfusion most often skipped even where the threshold is held. Checks whether the last unit actually worked — a rise under 0.5 g/dL usually means ongoing bleeding.
Technical detail
The one threshold the trials disagree on (acute coronary syndrome) is shown with that disagreement stated rather than presented as settled. 40 tests passing.Command centre(local, needs console.py running) Thresholds
34

Renal replacement therapy initiation timing

When to start dialysis in AKI — genuinely uncertain in the literature, which limits what any model can be trained against.

35

Drug dosing in organ dysfunction

Renal and hepatic adjustment computed continuously rather than from a table. The single most fatigue-sensitive task in the unit — recalculating doses as clearance changes, on every patient, every day.

For reviewBuilt · v0.1 Computes clearance, states the formulary rule at that band, flags what needs review — and does not prescribe. A produced number implies an authority the system does not have, and would be acted on. The finding that shaped the build: augmented renal clearance affects 30–65% of ICU patients and up to 85% in sepsis, and those patients are UNDER-dosed on standard regimens with documented treatment failure and resistance selection. Every dose tool watches for impairment; almost none watch for the opposite, and it is more common.
Technical detail
In acute injury the estimate carries its own caveat — creatinine lags true clearance, so it is optimistic exactly when it matters. 48 tests passing.Command centre(local, needs console.py running) Clearance bands & formulary rules
36

Drug interaction and contraindication checking

The ICU patient on eighteen drugs exceeds what anyone holds in working memory at hour eleven of a shift. Eighteen drugs is 153 pairwise comparisons, recomputed every time the chart changes. The clearest case in this register for holding without decay — not judgement, not prediction, a lookup done completely.

For reviewBuilt · v0.1 Three checks. Pairs for specific mechanisms. Groups for additive risk — three QT-prolonging drugs is one finding about cumulative risk, which no pairwise table can express. Conditions for drug against patient state, which a drug-drug checker cannot see at all: an NSAID is only wrong because the kidneys are.
Technical detail
Deliberately curated rather than exhaustive, and only contraindicated and major interrupt — enoxaparin with aspirin is routine, and interrupting on it is how an interaction checker gets ignored. 40 tests passing.Command centre(local, needs console.py running) Interaction table
37

Nutrition target and delivery

Prescribed versus actually delivered. Feeding is interrupted constantly — procedures, extubation, imaging, residuals — and each interruption is individually reasonable and individually small. Nobody totals them, so a patient can receive half their calories for a week while every decision along the way was correct.

For reviewBuilt · v0.1 The ramp is the target — ESPEN specifies a progressive increase, so 25 kcal/kg on day 1 is over-feeding, and a tool that only says “feed more” is wrong for the first three days. Protein separately — measured delivery is 83% of calorie targets against 65% for protein, because a feed run at rate reaches calories first and the volume therefore looks adequate.
Technical detail
Three things a simpler engine gets wrong. Refeeding inverts everything — the one place slow delivery is correct and the engine must not push. 47 tests passing.Command centre(local, needs console.py running) Targets & ramp
38

VTE prophylaxis

Balancing clot risk against bleeding risk, per patient, every day. Almost every ICU patient qualifies, so it becomes automatic — and automatic means nobody re-examines it.

For reviewBuilt · v0.1 Three failures, all invisible on one day and obvious across a week: never started, held for a procedure and forgotten, or continuing through a falling platelet count. The one it exists for is the third — a patient with a bleeding contraindication very often receives nothing at all, when mechanical prophylaxis is the answer.
Technical detail
A contraindication is a reason to switch, not to do nothing. 38 tests passing.Command centre(local, needs console.py running) Risk & contraindications

Reasoning, teaching, and the knowledge base

39–44

The category that addresses the stated national problem — a shortage of trained intensivists rather than of equipment. Education sits outside device regulation, which makes it the fastest route to real use.

39

Guideline deviation capture

Every engine here compares a patient against a published rule. This notices when practice departs from it and asks why — building a corpus of defensible deviation rather than policing it.

For reviewBuilt · v0.1 The detection is trivial; the value is the answer. A senior intensivist deviates for reasons rarely written down anywhere, and that reasoning does not survive being asked for in the abstract — asked against a concrete decision it surfaces. The framing decides whether anyone answers: it asks what were you seeing, not why you broke a rule. Presented as an audit it gets gamed within a fortnight; clinicians will explain reasoning and will not justify themselves to software.
Technical detail
A rule deviated from repeatedly with consistent reasoning is a finding about the guideline, not the clinicians. 35 tests passing.
40

Protocol and bundle adherence checking

Deterministic verification against the unit's own standing protocols. Boring, auditable, and it works. The ABCDEF bundle carries a hazard ratio of 0.32 for seven-day death across more than 15,000 adults — better evidence than any predictive model here — while complete delivery runs at 0–1%. The gap is delivery, not knowledge.

For reviewBuilt · v0.1 Unit tracker: what is due, for whom, what was missed, and how the unit compares to published delivery rates. An element counts only when every requirement is met — element B needs both the awakening and the breathing trial.
Technical detail
Failed safety screens are excluded rather than counted as neglect. 21 tests passing. Bundle tracker(local, needs console.py running) Bundle definition
41

Procedural guidance

Step logic and failure modes for the procedure a district physician performs twice a year. A textbook describes the procedure; what infrequency removes is the failure modes that experience supplies.

For reviewBuilt · v0.1 Central line, intubation, chest drain. Preparation is a gate, not a list — stopping midway to find something is where harm happens. Every procedure names its one commonly skipped step, marked inline rather than in a footnote, because a warning read beforehand is forgotten by step six: letting chlorhexidine dry, saying the failure plan out loud, confirming the side against the image. Rescue is searchable mid-procedure.
Technical detail
It states plainly that it does not make an untrained operator safe. 50 tests passing.Procedures
42

Decision frameworks extracted from the standard texts

A chapter is thirty pages because it must cover everything. What is needed at three in the morning is the decision sequence — what to exclude first, what changes the branch, where the exception lies. That sequence is perhaps fifteen lines, distributed across several chapters and several authors, which is why nobody can hold it.

For reviewBuilt · shell Every node states what to ask, why that order, what each answer changes, and where the sequence stops — and cites the page it came from, so any claim can be walked back. Carries a gap report naming where the sources disagree or assume, which is the working list rather than an apology. Why not chunked retrieval: clinical reasoning is long-form and conditional; chunking severs the sequence and returns a fragment that reads authoritative while having lost the preconditions that made it correct. Structure is extracted first; retrieval runs over the structure. The frameworks themselves are content and must be authored from a named source and approved by a named clinician — the example carries verified_by: null on screen.
Technical detail
The extraction schema and the viewer, with one worked framework to show the shape.Open the shell
43

The derived parameters, and what each one decides

Twenty-five measures that are arithmetic on values a unit already records. None needs a model, a sensor nobody owns, or a trial. What each buys is a decision the raw numbers do not support on their own.

For reviewBuilt · reference Each parameter states how to compute it, the decision it supports, where it misleads, and what to do when it is abnormal. The third of those carries equal visual weight to the formula, because a trainee handed a formula will apply it everywhere: shock index stays normal under beta blockade, S/F ratio saturates above 97%, driving pressure is meaningless on a triggering patient, and half-life in health is not half-life in critical illness. The fourth column is a discipline on the list itself — a parameter that changes no action is a number, not a measurement. Three are marked as teaching most: S/F ratio, because it makes P/F reasoning available without an arterial line; driving pressure, because it is arithmetic a trainee can do in their head that changes what they set today; and half-life multiples elapsed, because it converts “the drug is not working” into “I am asking too early”.
Technical detail
Implementation notes and source.Open the reference
44

Scenario simulator — the situations too rare to accumulate

A trainee may finish three years without independently managing a cannot-intubate-cannot-oxygenate, a saddle embolus, or a tension pneumothorax on a ventilated patient. Then one night they do. Simulators do not teach the routine; they teach what is rare and unforgiving.

For reviewBuilt · shell Sixty simulated minutes compressed into ninety real seconds, and the clock does not wait — the patient changes whether or not the trainee acts, which is the difference between a simulator and a case discussion. The debrief is the product rather than the scenario: every rule is checked against what was done and when, with harmful actions named and explained. Two modes run the same scenario with and without the derived layer, which turns the simulator into a test of whether the dashboards change what people do rather than a claim that they help. Scenario files are content and must be authored by a clinician; the worked example says so on screen.
Technical detail
Implementation notes and source.Open the simulator

Extending scarce expertise

45–49

The capacity multiplier. Tele-ICU already operates in India at scale; the open question is what AI adds to it, and the answer is usually triage of the specialist's attention.

45

Tele-ICU attention triage

Which of many beds the one available intensivist looks at now. The capacity multiplier is the ordering, not the intelligence — and ISCCM endorses telecommunication where local expertise is lacking (100% consensus).

For reviewBuilt · v0.1 The ranking is a stated rule a panel can read and change, and every row shows its own arithmetic — an opaque queue is not trusted, and an untrusted queue is ignored in favour of walking the unit, which is the thing tele-ICU exists to avoid. Capacity is stated honestly: a remote intensivist has a fixed number of minutes and the queue is longer, so the view draws the line where the shift runs out and counts the critical beds below it.
Technical detail
Implying full coverage would be false. 32 tests passing.Tele-ICU queue(local, needs console.py running)
46

Structured case presentation

Assembles the picture before the specialist joins, so the call opens with the question rather than the history.

For reviewBuilt · v0.1 What is wrong, what is running, what has changed, and — the part that matters — what is not known. A remote specialist cannot see the chart, so a summary that silently omits what was never measured is worse than one that says so. Unanswered weaning questions, unassessed extubation factors and a missing height all appear by name. A consultation that begins “tell me about this patient” wastes the resource it exists to conserve.
Technical detail
Implementation notes and source.Tele-ICU queue(local, needs console.py running)
47

Asynchronous specialist review queues

Not every question needs a live intensivist. Batching the non-urgent ones changes the staffing arithmetic entirely.

Assessed · held Needs a clinical view on what is genuinely non-urgent. The queue mechanics are the same as the tele-ICU triage already built; the clinical filter is a clinical decision.
48

Offline-capable decision support

Works without connectivity, because the district hospital's link drops. A deployment constraint that eliminates most cloud-dependent designs.

For reviewBuilt · proven The stated limit: this proves the reasoning layer survives a link drop, not that a deployment does — if the data feed is a cloud EHR, the unit still loses its data. What survives is reasoning over locally-held data.
Technical detail
Not an engine — a proof. The socket layer is replaced with one that raises on any outbound connection, then all 45 modules, 22 configs and a full unit assessment are run against it. No module imports a network client.
49

Nurse-led protocol execution support

Enables a trained nurse to run a protocol safely where no intensivist exists. Changes who can staff a unit.

For reviewBuilt · v0.1 A stop produces an escalation naming who to call and what to say, and is worded as the protocol working rather than as a failure.
Technical detail
Four delegated protocols with stop conditions that are structural, not advisory: once a stop is met there is no code path that returns a next step. Verified across all 54 combinations of protocol, stop condition and position.

The work of the unit

50–59

The most underrated category. None of it touches a clinical decision, so the regulatory burden is minimal — and it returns physician hours, which is the actual scarcity.

50

Ambient documentation from ward round speech

The note writes itself from the conversation that already happened. Returns hours per clinician per day with no clinical risk.

Assessed · held Needs an LLM for the summarisation layer, which sits outside the stated deterministic scope. The structured-capture half is buildable now if wanted.
51

Automated shift handover summarisation

Handover is where information is lost most reliably. Structuring it addresses a known and repeated failure point.

For reviewBuilt · v0.1 Carries the three things that actually get lost — the unclosed item, the trajectory, and the outgoing nurse’s own concern. “Stable” is never emitted without its support qualifier, because stable-on-rising-support is the phrase that reads as good news and means the opposite. The watch item is never machine-written; an empty one is reported.
Technical detail
Assembled from engine state rather than retyped.
52

Discharge and transfer summary generation

Tedious, delayed, and consequential for what happens next to the patient.

53

Nursing workload prediction and shift allocation

Staffing to predicted acuity rather than to bed count.

For reviewBuilt · v0.1 Demand computed from what is being done to each patient, not from the bed count. NIV costs more nursing time than invasive ventilation; end-of-life care is the heaviest single item; agitation is among the top three and is the one workload tools omit because it generates no procedure record. Scores patients and never nurses — a per-nurse score becomes a performance metric and then people under-record. Where demand exceeds the roster it names what gets deferred, and every item on that list is preventive.
Technical detail
Implementation notes and source.Unit operationsWeights
54

Bed and admission flow forecasting

Anticipates the crunch far enough ahead to act on it.

55

Automated infection surveillance and reporting

Line infections and VAP counted continuously rather than by retrospective chart audit.

For reviewBuilt · v0.1 Counted continuously rather than by retrospective audit, which undercounts hardest in the busiest months. A rate never appears without the numerator and denominator that produced it, and is withheld entirely below 50 device-days. Not-applicable is never collapsed into not-met — a VAE with no stable baseline cannot be assessed, and recording that as a negative is how an unassessable patient becomes a clean number. Every definition states what it is known to get wrong.
Technical detail
Implementation notes and source.Unit operationsDefinitions
56

Supply, oxygen, and consumable forecasting

India's oxygen crisis made this concrete. Demand prediction is a solved problem applied to an unsolved one.

For reviewBuilt · v0.1 Consumption is arithmetic, not prediction, so no model is used. The headline is hours until the order must be placed — supply minus delivery lead time — because 2021 was not a failure to know consumption. HFNC is costed at flow × FiO2, so one patient at 60 L/min outweighs four ventilated ones.
Technical detail
Admissions are named as the uncertainty and are not modelled; scenarios are attributed to the unit, not the engine.Unit operationsCoefficients
57

Equipment failure prediction

Predictive maintenance on ventilators and pumps. Matters most where a failed machine cannot be replaced the same day.

For reviewBuilt · v0.1 Run-hours against service interval, reporting the part that degrades first rather than the machine. A drifting oxygen sensor is more dangerous than a dead one because the delivered oxygen is wrong and nothing alarms. The fleet sorts by single points of failure: a spare pump is a nuisance, a failed CRRT machine with no second one stops the treatment.
Technical detail
Implementation notes and source.Unit operationsIntervals
58

Automated registry and quality-metric submission

Removes the manual burden that makes national registries incomplete — and incomplete registries are why India lacks its own outcome data.

Assessed · held Needs the national registry that does not yet exist, and its submission schema. Worth revisiting if a national registry advances.
59

Mortality and morbidity review preparation

Assembles the timeline so the meeting discusses the case rather than reconstructing it.

For reviewBuilt · v0.1 Assembles and orders; it does not adjudicate. No preventability field exists anywhere in the code, because software that suggests a death was avoidable turns the meeting into an argument with the tool and stops people recording candidly. The silences are the finding — unactioned results, observation gaps, deterioration-to-escalation intervals — and each is phrased as a question, with a note that absence from the record is not absence of care.
Technical detail
Implementation notes and source.Structure

The patient as a person

60–68

Consistently underfunded relative to its effect on outcomes that patients and families actually care about.

60

Delirium risk prediction and prevention prompting

Delirium drives long-term cognitive harm and is largely preventable through measures nobody remembers to apply consistently.

For reviewBuilt · v0.1 Risk split into modifiable and fixed, because a score nobody can act on is a number rather than a prevention.
Technical detail
Built together with 60, 61 and 62 in one engine: sedation drives delirium, drives immobility, drives skin breakdown. Kept separate they would recommend a sedative for agitation caused by the immobility they separately flag — the engine names that contradiction explicitly and resolves it toward lightening sedation.Unit operationsFactors
61

Delirium detection from behaviour and speech

Continuous detection rather than a twice-daily screening tool that is frequently skipped.

Assessed · held Needs behavioural and speech data with CAM-ICU labels. The risk-factor half is already built as entry 57.
62

Pain assessment in the non-communicative patient

Facial and physiological inference where the patient cannot report. An area of real and persistent under-treatment.

Assessed · held Needs facial video with expert pain scoring. Physiological inference alone is too weak to act on.
63

Sleep and circadian disruption monitoring

ICU sleep architecture is destroyed, and it feeds directly back into delirium.

For reviewBuilt · v0.1 Measured from the unit’s own record of what it did overnight, because the commonest cause of ICU sleep loss is the ICU. Interruptions are classed by how avoidable they are — alarms and 04:00 bloods are often avoidable; turning can be clustered but not skipped. Labelled as a proxy: it counts interruptions, not sleep.
Technical detail
Implementation notes and source.Unit operationsConfig
64

Early mobilisation readiness and prompting

Halves delirium in trial settings and is skipped constantly under staffing pressure.

For reviewBuilt · v0.1 Five levels with stated safety criteria, so a “no” carries its reason. Level 1 — passive range of motion — requires no criteria at all: a patient failing every gate is still a candidate, and reporting “not suitable for mobilisation” without saying so is the failure this structure prevents.
Technical detail
Implementation notes and source.Unit operationsCriteria
65

Pressure injury risk and repositioning prompting

Preventable, measured, and a standing quality metric everywhere.

For reviewBuilt · v0.1 Risk plus the clock, because risk alone changes nothing — it is elapsed time since the last turn that produces the injury, and high risk tightens the interval. Vasopressors count heavily: the treatment itself reduces perfusion, so skin fails faster than the standard scales assume. Device sites — tube ties, mask straps, oximeter probes — are inspected separately, since nobody thinks of a tube tie as a pressure source.
Technical detail
Implementation notes and source.Unit operationsConfig
66

Family communication support and update generation

Families are told too little, too late, in language they do not follow. Also a live medico-legal exposure in India.

Assessed · held Needs a settled position on what families are told and by whom — a medico-legal question in India before it is a technical one.
67

Communication aids for the ventilated patient

Gaze or gesture decoding for someone who cannot speak. Small population, enormous individual effect.

Assessed · held Needs eye-tracking or gesture hardware at the bedside. Small population, large individual effect, and entirely dependent on a device budget.
68

Post-intensive-care syndrome risk identification

Who will carry cognitive and physical harm home. Barely addressed anywhere, and invisible in mortality statistics.

For reviewBuilt · v0.1 Three domains rather than one score, because cognitive, physical and psychological harm have different predictors and lead to three different follow-up arrangements. Delirium duration is the strongest cognitive predictor; delusional memories are the strongest psychological one and the thing nobody asks about. Identifies who to follow up — explicitly not a prognosis to give a patient. Family harm is named as deliberately out of scope.
Technical detail
Implementation notes and source.Domains

What has been built

35 of 68

Working code in this repository, marked against the register above. Each built entry links to its console page, its editable config, its tests and its source from within the register — this is an index, not a second copy.

Nothing is deployed. Nothing has touched a patient, no configuration has been verified by a clinician, and every rule is in draft with no approving panel named.

01Continuous sepsis monitoring
02Acute kidney injury — staging and nephrotoxin review
03Trajectory — direction and rate, not the current value
06Patient–ventilator asynchrony detection
07Weaning readiness screen
08Extubation failure risk
10Lung-protective ventilation adherence
31Antimicrobial de-escalation
32Sedation depth and daily interruption
33Transfusion thresholds
35Drug dosing in organ dysfunction
36Drug interaction and contraindication checking
37Nutrition target and delivery
38VTE prophylaxis
39Guideline deviation capture
40Protocol and bundle adherence checking
41Procedural guidance
42Decision frameworks extracted from the standard texts
43The derived parameters, and what each one decides
44Scenario simulator — the situations too rare to accumulate
45Tele-ICU attention triage
46Structured case presentation
48Offline-capable decision support
49Nurse-led protocol execution support
51Automated shift handover summarisation
53Nursing workload prediction and shift allocation
55Automated infection surveillance and reporting
56Supply, oxygen, and consumable forecasting
57Equipment failure prediction
59Mortality and morbidity review preparation
60Delirium risk prediction and prevention prompting
63Sleep and circadian disruption monitoring
64Early mobilisation readiness and prompting
65Pressure injury risk and repositioning prompting
68Post-intensive-care syndrome risk identification

Awaiting data

12

The method is settled; the input is not in hand. These are model tasks — imaging, waveform, and video — held until there is a labelled corpus from the units they would run in. Trained on Indian data, not imported: a model built elsewhere and dropped in is the failure this register documents repeatedly.

04Deterioration detection from camera alone
16Chest radiograph triage and abnormality flagging
17Endotracheal tube and line position verification
18Head CT haemorrhage detection and triage
19Pneumothorax detection on portable films
20Lung ultrasound interpretation assistance
21Automated echocardiography for the non-cardiologist
22Guided ultrasound acquisition
23Continuous ECG interpretation beyond arrhythmia
24Continuous EEG interpretation for non-convulsive seizure
25Pupillometry and automated neurological assessment
26Non-contact vital sign estimation from video

21 described but not built — these are the field as it stands, not work done here. They carry no build marker.

Illustration and research prototype
© WyteRock · wyterock.com