Towards Operationalisable Clinical Risk Prediction Models
AKI is the current clinical demonstrator. This PhD asks how a risk model can move from a valid
prospective target, to a faithful representation of the evolving patient and clinical context, to an
operationally meaningful signal that supports structured review without pretending that prognosis is the
same as preventability or treatment response.
Current repository authority: HOLDAKI = current demonstratorContext = haemodynamic · infection · medication · procedureRisk score → structured context review, not treatment
Executive summary
One briefing, with a fast supervisor-level overview and deeper epidemiology / phenotype detail underneath.
Clinical problem
Future severe AKI
Predict clinically meaningful deterioration while preserving prospective information boundaries.
accepted framing
Intended decision
Prioritise structured review
The treating ICU / acute-care team reviews prospectively observable context and routes onward when warranted.
DEC-0084
Primary outcome
First prospectively ascertainable Stage ≥2
Persistent / severe trajectory outcomes remain secondary; the operational endpoint is observation-process dependent.
DEC-0087
Current authority
No globally accepted target, features or model
Candidate.8 is canonical and non-executable; patient-level target/model execution remains gated.
HOLD
Working research question
In adults at repeated ICU-origin prediction times, can prospectively observable clinical context improve the clinical meaning and operational usefulness of future severe-AKI risk prediction beyond a strong time-updated physiology baseline, and which contextual representations provide stable incremental value without compromising prospective validity?
Research framework
The thesis is organised around three linked methodological problems. Operationalisability is evaluated throughout rather than added as a final deployment chapter.
1 · Define
Define the prediction problem
Translate a clinical construct into a prospectively valid EHR prediction target with explicit risk-set, timing, component, missingness and terminal-event semantics.
Main AKI issue
Clinical AKI ≠ EHR-ascertainable AKI ≠ operational prediction endpoint.
→
2 · Represent
Represent the evolving patient and context
Move beyond a flat list of predictors to dynamic renal, physiological, treatment, exposure and observation-process states.
Main AKI issue
Context changes the meaning of the same observed value or medicine.
→
3 · Justify complexity
Add external knowledge only for a demonstrated residual problem
Test conventional representations first. Knowledge-informed or neuro-symbolic methods enter only if they address a specific limitation that remains.
Research design principle
Method follows the clinical and methodological requirement, not novelty.
Current context priority (DEC-0087): observed haemodynamic / treatment-state context first, infection / acute deterioration second, medication / pharmacology third, and procedure / surgery / contrast fourth. Observation process is a cross-cutting validity layer rather than a later competing context block. Each added context still requires its own evidence, prospective-observability and specification gate.
Epidemiological & prediction design
The detailed review portal is folded into this briefing here: who enters the risk set, when prediction is issued, what event is forecast, how follow-up works, and where uncertainty remains.
Who
Incident-risk state
Current prospectively ascertainable Stage 0 or Stage 1 can remain eligible for first future Stage ≥2 prediction.
DEC-0004
When
Repeated ICU-origin occasions
Rolling entry opens from ICU +6 h. q6 is the accepted primary model risk-refresh cadence; q12 is the prespecified cadence sensitivity.
DEC-0011 / 0012
What
First Stage ≥2
Forecast the first prospectively ascertainable KDIGO Stage ≥2 event; persistent / severe trajectory remains secondary.
DEC-0003 / 0006 / 0014
Horizon
24 h primary
The primary horizon is (t, t+24 h]; 48 h is secondary. A prespecified sensitivity excludes events in the first 6 h after prediction.
DEC-0013 / 0025
Follow-up
Hospital-wide issued horizon
Leaving the index ICU stops new prediction issuance, but an already-issued horizon continues under hospital follow-up. Death and discharge remain distinct terminal states.
DEC-0015 / 0016 / 0023
Figure 1
Prospective rolling-prediction design
Accepted partial design decisions within the current prospective AKI target framework.
A · Rolling prediction schedule
ICU admissionIndex ICU begins
→
+6 hEarliest possible prediction entry
→
+12 hscheduled occasion
→
+18 hscheduled occasion
→
+24 h · +30 h · …q6 h while still in index ICU
Eligibility at time t: issue a prediction when current prospectively ascertainable renal state is Stage 0 or Stage 1. If UNKNOWN, defer and reassess. If Stage ≥2 is already ascertainable, do not enter the first-future-Stage ≥2 risk set.
Model refresh: q6 h primary; q12 h prespecified cadence sensitivity. This is model-computation cadence, not alert frequency.
B · Prediction horizon
Prediction time tprospectively valid information only
→
Primary: (t, t+24 h]first future Stage ≥2
→
Secondary: (t, t+48 h]early-warning horizon
→
0–6 h sensitivityexclude very near-term events
C · ICU transfer and terminal states
Index ICUnew q6 predictions may be issued
→
ICU-to-ward transfercare-setting transition
→
No new ward predictionsfor this current MIMIC target
Already-issued horizon continues: transfer does not censor an issued 24 h/48 h window; follow-up continues in hospital using prospectively observable components.
Terminal states: first Stage ≥2 exits the first-event risk set; death before Stage ≥2 is a competing terminal event; hospital discharge alive ends in-hospital follow-up. Death/discharge are not ordinary negatives.
Project-status boundary: these are accepted partial design decisions. The complete executable target remains unaccepted and unexecuted under current project authority.
Figure 1. Working prospective rolling-prediction architecture. It distinguishes accepted entry, cadence, horizon and follow-up decisions from the still-gated complete executable target.
Figure 2
Development ICU-stay substrate and q6 prediction-occasion structure
Confirmed current UO/q6 development counts; this is not the final accepted AKI modelling cohort.
Interpretation: 931,427 is the complete scheduled q6 development grid, not the final number of target-eligible prediction observations. Final outcome/risk-set filtering requires the accepted executable target.
Figure 2. Development denominator hierarchy from the corrected UO/q6 substrate QC: patients → admissions → ICU stays → q6-contributing stays → repeated stay-time prediction occasions.
Why this is not just “predict AKI in the next 24 hours”
The phenotype must distinguish clinical KDIGO criteria from what was actually observable in the EHR at the prediction time. Baseline SCr availability, UO evaluability, KRT attribution, event-time versus availability-time, revisions, terminal events and zero-opportunity outcome ascertainment can all change what the label means.
What remains non-authoritative / gated
The complete target is not globally accepted in CURRENT_AUTHORITY.yaml; candidate.8 remains canonical, aligned and non-executable. Global feature and model fields are also null. The page therefore distinguishes accepted partial decisions from candidate/specification state.
Clinical phenotypes & relationships
This replaces a standalone “Medication-aware design” section. Medication is one clinically important context domain within a wider phenotype system.
Future severe AKIFirst prospectively ascertainable Stage ≥2 · operational EHR endpoint
Selected domain
Baseline vulnerability
Background susceptibility can alter future AKI risk without being the acute mechanism itself.
Chronic kidney disease / reduced renal reserve
History of prior AKI
Age ≥65 and selected chronic disease contexts
Relationship / guardrail
CKD → risk factor for AKI. Keep chronic baseline state separate from acute change; one elevated creatinine does not establish CKD.
C16 / C17context ≠ cause
Oliguriacriterion forAKI
Adequately measured low UO can establish AKI stage; missing UO cannot be treated as normal UO.
Hypotension≠Hypovolaemia
Low MAP/BP is haemodynamic information, not a deterministic label for intravascular depletion.
Medication exposure before AKI≠Drug-induced AKI
Temporal precedence is a patient fact, not causal attribution or preventability.
UNKNOWN component≠Observed negative
Absence of a valid assessment opportunity must remain explicit rather than silently becoming a negative label.
Hypovolaemiamodifies actionDiuretic
Depletion may support withholding/review, while congestion may make diuretic treatment appropriate: the same medicine can imply opposite actions.
EHR-ascertainable AKI≠Clinical / latent AKI truth
The operational phenotype depends on measurement coverage, timestamps and observation policy.
A value is usable only when its clinical event and its information availability are valid for the prediction occasion. This affects both the target and predictor-side context.
1 · Clinical event
A test, observation, drug administration, procedure or treatment state occurs.
2 · Information becomes available
The EHR can expose the information later than the biological or collection event.
3 · Prediction occasion t
Only information validly available by t may enter the model or current-state phenotype.
4 · Future window (t, t+H]
The outcome is sought prospectively after the prediction time.
5 · Ascertainability
Observed non-event is separated from zero valid opportunity to determine the endpoint.
Renal component
SCr, UO and KRT are non-equivalent signals
Each has different observation and timing properties. One observed qualifying component may establish AKI while another remains unavailable.
Context component
Observed proxy ≠ latent state
MAP, lactate, fluids and vasopressors may support haemodynamic interpretation but do not by themselves establish shock, hypovolaemia or a renal mechanism.
Transportability
Measurement policy can change model meaning
An EHR phenotype and its apparent missingness can shift when another hospital measures or documents differently.
How clinical context enters the prediction model
The modelling framework is not medication-specific. Context domains are added one at a time, with each block requiring its own clinical rationale, prospective-observability check and specification before comparison.
reference model
A · Time-updated patient state
Start with a strong renal and physiological representation using only prospectively valid observed patient information.
↓
Question: what can the evolving clinical state already explain?
context increment
B · Add one governed context domain
Add one clinically justified context block. Current priority is observed haemodynamic/treatment state, then infection/acute deterioration, then medication/pharmacology, then procedure/surgery/contrast.
↓
Each domain is tested separately before combining domains, so any incremental value remains interpretable.
contextual representation
C · Context × patient state × time
Test whether the same context becomes more informative when represented relative to renal trajectory, physiology, treatment state and time rather than as a flat feature list.
↓
Only after fair same-information comparisons should more complex or external-knowledge mechanisms be considered.
Current first context priority: observed haemodynamic / treatment-state context.
This is the first broader-context study to develop next, based on current feasibility, local support and clinical relevance. The existing medication A → B_RAW → B_CONTEXT study remains a governed bounded workstream, but it is no longer assumed to come first in the PhD sequence.
priority 1 Observed haemodynamic / treatment state BP/MAP trajectories, governed support state and other prospectively observed haemodynamic information; do not infer hypovolaemia or shock from one proxy.
priority 2 Infection / acute deterioration Sepsis-related and evolving acuity context, with phenotype semantics kept separate from individual treatment or laboratory signals.
priority 3 Medication / pharmacology Prospective exposure, dose/timing and medication × renal/physiological context; the bounded medication study remains valid but moves later in sequence.
priority 4 Procedure / surgery / contrast Prospectively timed procedural and exposure context where clinically justified and locally supported.
cross-cutting validity Observation process Measurement recency, availability, coverage and UNKNOWN state accompany every model comparison rather than competing as a later context block.
later after validation Latent volume / congestion / cause-specific states Potentially high impact, but only after transparent phenotype validation; low MAP ≠ hypovolaemia and heart failure ≠ congestion.
Anti-conflation rule: a clinically valid relationship does not automatically become a predictor, a predictive interaction, a model constraint, or a treatment recommendation.
Evaluation: prediction + operational usefulness
The accepted evaluation architecture avoids reducing success to AUROC alone and keeps retrospective predictive claims separate from clinical-outcome benefit.
24 h
Primary horizon; 48 h retained as secondary/sensitivity.
ΔAUPRC
Primary discrimination estimand for the central contextual representation contrast; paired ΔAUROC is key secondary evidence.
Calibration
Reliability curve, calibration-in-the-large/intercept where compatible, slope and Brier score.
Matched burden
Compare PPV, sensitivity and workload at prespecified review-burden operating points rather than inventing an “optimal” cutoff.
First alert
Patient/stay-level first-alert PPV, sensitivity, lead time, unique patients alerted and repeated-alert burden.
2,000×
Paired subject-cluster bootstrap replicates for final uncertainty, preserving within-subject repeated observations.
Claim boundary: retrospective work may establish predictive increment and, with supporting operational evidence, an operationally promising increment. It cannot establish treatment benefit or clinical utility without prospective/interventional evaluation.
Research sequence
A gate-based plan is more defensible than promising a fixed number of papers or methods in advance.
Now
Close the target / phenotype path
Finish the remaining executable integration and validation gates for the prospectively ascertainable Stage ≥2 phenotype.
Next
Run strong simple comparators
Establish time-updated physiology performance and the first governed medication/context increments with prespecified evaluation.
Then
Expand context deliberately
Prioritise later phenotype/context blocks only after evidence, observability and specification review.
Conditional
Knowledge-informed methods
Introduce explicit external knowledge or neuro-symbolic mechanisms only when ordinary learning leaves a concrete residual problem.
Useful supervisor discussion points
These are discussion prompts, not hidden requests to reopen already accepted project decisions.
1
Paper / chapter distinctness
Is prospective phenotype and ascertainment methodology sufficiently substantial as the first empirical chapter, or should it be framed primarily as enabling methodology for the later context study?
2
First context-study design
For the haemodynamic / treatment-state study, which observed representation should be the main incremental test: continuous MAP/BP trajectories, support-state transitions, lactate/perfusion summaries, or a deliberately minimal combination?
3
Method-complexity threshold
What empirical failure of simpler temporal/context models would be sufficient justification for TraCeR-style, Transformer or neuro-symbolic complexity?
4
Operational evidence expected for the thesis
How far should the PhD push beyond discrimination/calibration into first-alert burden, lead time, transportability and workflow-facing evaluation before prospective clinical validation becomes future work?
Evidence trail for this version
The page separates live project authority, the Evidence Matrix, project literature notes and synthesis.
Live project governance checked first
Source access: Project governance files are stored in a private GitHub repository and require authorised access.
AKI research-question human gate — historical DEC-0084/0085 intended-use/research framing; its medication-first D2 ordering is superseded.
AKI context-priority decision — DEC-0087 replaces medication-first sequencing with haemodynamic → infection → medication → procedure, while preserving D3/D4 and the HOLD boundary.
Consulted 00_Read Me, C16_Clinical Phenotypes, C17_Phenotype Relationships and P19_Feature–Phenotype Crosswalk. The Matrix is an evidence-navigation/synthesis layer, not scientific authority.
REV-PHENO-01 — AKI operational definition, prospective EHR ascertainment and component-availability distinctions.
REV-ASCERT-01 — outcome unascertainability, informative observation and uncertain endpoints.
REV-INTUSE-07 — intended-use synthesis and clinical ownership / action architecture.
REV-MED-08 — prediction relevance and system placement of medication knowledge; action knowledge ≠ prediction knowledge.
Private reflective sources were not consulted. The phenotype relationship map is a supervisor-facing synthesis of the live Matrix and governed project evidence; it is not an executable ontology, feature contract or causal model.