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.
One briefing, with a fast supervisor-level overview and deeper epidemiology / phenotype detail underneath.
Predict clinically meaningful deterioration while preserving prospective information boundaries.
The treating ICU / acute-care team reviews prospectively observable context and routes onward when warranted.
Persistent / severe trajectory outcomes remain secondary; the operational endpoint is observation-process dependent.
Candidate.8 is canonical and non-executable; patient-level target/model execution remains gated.
The thesis is organised around three linked methodological problems. Operationalisability is evaluated throughout rather than added as a final deployment chapter.
Translate a clinical construct into a prospectively valid EHR prediction target with explicit risk-set, timing, component, missingness and terminal-event semantics.
Main AKI issueClinical AKI ≠ EHR-ascertainable AKI ≠ operational prediction endpoint.
Move beyond a flat list of predictors to dynamic renal, physiological, treatment, exposure and observation-process states.
Main AKI issueContext changes the meaning of the same observed value or medicine.
Test conventional representations first. Knowledge-informed or neuro-symbolic methods enter only if they address a specific limitation that remains.
Research design principleMethod follows the clinical and methodological requirement, not novelty.
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.
Current prospectively ascertainable Stage 0 or Stage 1 can remain eligible for first future Stage ≥2 prediction.
Rolling entry opens from ICU +6 h. q6 is the accepted primary model risk-refresh cadence; q12 is the prespecified cadence sensitivity.
Forecast the first prospectively ascertainable KDIGO Stage ≥2 event; persistent / severe trajectory remains secondary.
The primary horizon is (t, t+24 h]; 48 h is secondary. A prespecified sensitivity excludes events in the first 6 h after prediction.
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.
This makes the reasoning auditable: external literature defines the methodological problem, project epidemiology quantifies the trade-off in MIMIC, and a separate human decision records the selected design and what remains uncertain.
| Design question | Evidence examined | What the evidence showed | Current decision | Remaining uncertainty |
|---|---|---|---|---|
| Who enters the risk set? | EPID-001, EPID-002 + phenotype literature |
Stage 1 remained a meaningful at-risk state. At q6/24 h, Stage 0/1 eligibility captured 2,669 additional unique future Stage ≥2 events versus Stage 0 alone. | Current prospectively ascertainable Stage 0 or Stage 1 is eligible for first future Stage ≥2 prediction. DEC-0004 |
Eligibility still depends on the final prospectively available-component phenotype. |
| When should prediction begin? | EPID-006 + temporal/workflow review |
Governed prospective renal states can already be available at +6 h. A fixed +24 h landmark would impose a later entry point than the reviewed data require. | Rolling eligibility opens from ICU +6 h; UNKNOWN is deferred rather than forced into Stage 0/1. DEC-0011 |
Not every patient is classifiable at +6 h; the final component set can alter early classifiability. |
| How often should risk update? | EPID-002 + workflow literature |
q6 h is the highest-resolution reviewed rolling substrate. Model computation cadence is conceptually separate from clinical alert/review cadence. | q6 h primary; q12 h prespecified sensitivity / efficiency analysis. DEC-0012 |
Whether q12 h gives similar practical performance with lower repeated-prediction burden remains empirical. |
| 24 h or 48 h horizon? | Rolling target-design audit + warning-time review | 24 h captured 5,796 unique Stage ≥2 events versus 6,223 at 48 h (93.1% of 48 h coverage). Median warning was 11.1 h versus 19.7 h; 24 h also had fewer after-transfer positives and terminally shortened windows. | 24 h primary; 48 h secondary. Retain a 0–6 h minimum-warning-time sensitivity. DEC-0013 |
The clinical value of extra warning versus tighter near-term prediction remains an explicit trade-off. |
| What counts as Stage ≥2? | REV-PHENO-01, KDIGO / operationalisation literature + MIMIC phenotype evidence |
SCr, UO and KRT have different observation processes. An observed qualifying component can establish AKI; an unavailable component cannot be treated as criterion-negative. | First prospectively EHR-ascertainable available-component KDIGO Stage ≥2; hospital-wide SCr Stage ≥2 retained as a major sensitivity / benchmark. DEC-0014 |
The complete executable target remains gated; component implementation and observation adequacy still require governed closure. |
| What happens at ICU transfer? | EPID-002 + ascertainment / workflow literature |
6.8% of q6/24 h Stage ≥2 positive windows occurred after ICU transfer. Transfer changes the observation regime but does not biologically end AKI risk. | An already-issued horizon continues hospital-wide; no new ward predictions are created for the current MIMIC design. DEC-0015, DEC-0023 |
A future ward-compatible prediction model would require its own feature and prospective-availability contract. |
| What about death and discharge? | EPID-002 + REV-ASCERT-01 / terminal-event reasoning |
Death and hospital discharge alive terminate the opportunity for the explicitly in-hospital endpoint and should not be collapsed into ordinary “no AKI” labels. | Preserve death-before-Stage ≥2 and hospital-discharge-before-Stage ≥2 as distinct terminal states. DEC-0016 |
The later statistical model still has to specify how these terminal states enter learning and evaluation. |
| What exact future-time interval? | Prospective prediction semantics + reproducible target-clock requirements | An event already ascertainable at prediction time t is not a future outcome; a deterministic right-endpoint convention is needed for reproducibility. | Use (t, t+H], with H = 24 h primary and 48 h secondary. DEC-0025 |
Main residual issue is source-specific timestamp precision / ordering, not the interval convention itself. |
CURRENT_AUTHORITY.yaml still records the project as HOLD with no globally accepted target, feature contract or model.
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.
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.
A brief update on where the work is now and what needs to finish before the next stage.
Most of the epidemiological study design is now in place. I am currently running a full check of the urine-output pipeline on the development dataset before incorporating it into the prediction framework.
The run is computationally slow and is still ongoing. I am not interpreting partial results. Once it finishes cleanly, I can assess whether the urine-output information is sufficiently available and reliable for the planned predictor set.
Background susceptibility can alter future AKI risk without being the acute mechanism itself.
CKD → risk factor for AKI. Keep chronic baseline state separate from acute change; one elevated creatinine does not establish CKD.
Adequately measured low UO can establish AKI stage; missing UO cannot be treated as normal UO.
Low MAP/BP is haemodynamic information, not a deterministic label for intravascular depletion.
Temporal precedence is a patient fact, not causal attribution or preventability.
Absence of a valid assessment opportunity must remain explicit rather than silently becoming a negative label.
Depletion may support withholding/review, while congestion may make diuretic treatment appropriate: the same medicine can imply opposite actions.
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.
A test, observation, drug administration, procedure or treatment state occurs.
The EHR can expose the information later than the biological or collection event.
Only information validly available by t may enter the model or current-state phenotype.
The outcome is sought prospectively after the prediction time.
Observed non-event is separated from zero valid opportunity to determine the endpoint.
Each has different observation and timing properties. One observed qualifying component may establish AKI while another remains unavailable.
MAP, lactate, fluids and vasopressors may support haemodynamic interpretation but do not by themselves establish shock, hypovolaemia or a renal mechanism.
An EHR phenotype and its apparent missingness can shift when another hospital measures or documents differently.
Context domains are added one at a time, with each block requiring its own clinical rationale, prospective-observability check and specification before comparison.
Start with a strong renal and physiological representation using only prospectively valid observed patient information.
Question: what can the evolving clinical state already explain?
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.
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.
The accepted evaluation architecture avoids reducing success to AUROC alone and keeps retrospective predictive claims separate from clinical-outcome benefit.
Primary horizon; 48 h retained as secondary/sensitivity.
Primary discrimination estimand for the central contextual representation contrast; paired ΔAUROC is key secondary evidence.
Reliability curve, calibration-in-the-large/intercept where compatible, slope and Brier score.
Compare PPV, sensitivity and workload at prespecified review-burden operating points rather than inventing an “optimal” cutoff.
Patient/stay-level first-alert PPV, sensitivity, lead time, unique patients alerted and repeated-alert burden.
Paired subject-cluster bootstrap replicates for final uncertainty, preserving within-subject repeated observations.
A gate-based plan is more defensible than promising a fixed number of papers or methods in advance.
Finish the remaining executable integration and validation gates for the prospectively ascertainable Stage ≥2 phenotype.
Establish time-updated physiology performance and the first governed haemodynamic / treatment-state context increment with prespecified evaluation.
Prioritise later phenotype/context blocks only after evidence, observability and specification review.
Introduce explicit external knowledge or neuro-symbolic mechanisms only when ordinary learning leaves a concrete residual problem.
A three-paper framing for supervisor discussion, not a fixed commitment. Paper 1 focuses on methodological epidemiology and target ascertainment; Paper 2 uses a TraCeR-style longitudinal competing-risk survival model as the candidate backbone for context-aware prediction; Paper 3 is a neuro-symbolic extension testing whether governed directional and relational knowledge adds incremental value. The model architectures remain candidate framing, not current project authority.
Methodological epidemiology · phenotype / estimand / clinical informatics
Candidate framing: how to define and evaluate a temporally valid severe-AKI prediction target when the clinical phenotype and its observation process evolve over time.
How should first future Stage ≥2 AKI be defined and ascertained prospectively when component availability, measurement timing, ICU transfer and terminal events vary over time?
A reproducible target-design framework that treats the observation process as part of the operational estimand rather than hidden preprocessing, with its epidemiological consequences quantified in longitudinal EHR data.
Methodological prediction + applied clinical epidemiology
Candidate framing: use a TraCeR-style longitudinal competing-risk survival Transformer as the candidate backbone for repeated prediction. The retrospective question is whether integrated context makes the risk representation more clinically meaningful while retaining strong predictive performance — not whether the model has already demonstrated clinical or operational utility.
Compared with a dynamic renal/physiology baseline, does adding integrated prospectively observable haemodynamic/treatment-state, infection/acute-deterioration, medication/pharmacology and procedure/exposure context preserve strong predictive validity while providing more clinically meaningful characterisation of future severe-AKI risk?
A context-aware longitudinal survival framework that tests whether clinically interpretable patient-state information can enrich severe-AKI risk characterisation without sacrificing predictive validity, while keeping retrospective predictive evidence distinct from claims of real-world clinical utility.
Neuro-symbolic extension of the Paper 2 survival model
Candidate framing: keep the Paper 2 target, cohort, observed patient information and TraCeR-style backbone fixed, then add explicit governed directional / relational knowledge through a NeSy / LTN-style extension so that the incremental contribution of knowledge itself can be tested.
Does injecting governed directional and relational clinical knowledge into the matched context-aware longitudinal survival model improve predictive validity, calibration, clinically meaningful risk characterisation or semantic consistency beyond the same observed patient information alone?
A controlled test of whether explicit clinical direction and relationship knowledge provides incremental value beyond a strong context-aware longitudinal survival model, separating the effect of knowledge injection from the effect of additional patient information.
The page separates live project authority, the Evidence Matrix, project literature notes and synthesis.
Source access: Project governance files are stored in a private GitHub repository and require authorised access.
HOLD; target/features/model globally unaccepted.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-03 — AKI alert / CDS intervention evidence and the limits of generic risk alerts.REV-INTUSE-05 — temporal workflow, warning-time and cadence considerations.REV-INTUSE-06 — MIMIC workflow observability and transportability constraints.REV-TRANS-03 — critical methodology comparison and the requirement that added complexity address a demonstrated problem.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.