AIIE Methodology & Evidence Basis
How the ARKA Imaging Intelligence Engine turns structured clinical context into a 1–9 appropriateness score — the six weighted factors, the signed-signal mapping, and the published evidence each factor draws on.
Overview & the 1–9 appropriateness scale
AIIE expresses imaging appropriateness on the RAND/UCLA 1–9 scale, where 1–3 is generally not appropriate, 4–6 is may be appropriate, and 7–9 is usually appropriate. Scoring begins from a neutral baseline of 5 and is nudged up or down by a small set of weighted clinical factors.
Each factor produces a signed signal in the range −1 to +1. A matched knowledge-matrix rating r on the same 1–9 scale is mapped to a signal by centering on the baseline and normalizing to the half-range:
signal = clamp((rating − 5) / 4, −1, +1)
A rating of 9 maps to +1 (strongly appropriate), 5 maps to 0 (neutral), and 1 maps to −1 (strongly inappropriate). Each factor's contribution is its weight × signal × 4, and the final score is round(5 + Σ contributions), clamped to the integer range 1–9. The six core weights below sum to 1.00; when a trauma-severity gate applies, it reserves a fixed share and the remaining factors are rescaled proportionally.
When optional ML refinement is active, per-feature contributions are attributed with SHAP (SHapley Additive exPlanations) per Lundberg & Lee, NeurIPS 2017.
Clinical indication strength
Weight 0.30 · 30%How well the documented symptoms, duration, and reason-for-exam concord with the requested study answering a real clinical question.
How AIIE uses it
GRADE summary-of-findings framing for symptom–test concordance; RAND/UCLA indication panels.
Prior imaging redundancy
Weight 0.20 · 20%Whether recent same-region or same-modality studies already answer the question, so a repeat order is unlikely to change management.
How AIIE uses it
Choosing Wisely / published appropriateness criteria on redundant imaging in stable presentations.
Red flag symptoms
Weight 0.20 · 20%Presence of escalation features — e.g. cauda equina, progressive neurologic deficit, fever with focal pain, concerning trauma — that warrant expedited imaging.
How AIIE uses it
ACEP / specialty red-flag guidance for urgent diagnoses requiring expedited imaging.
Guideline alignment
Weight 0.15 · 15%How closely the order maps to the matched scenario/variant in the AIIE Clinical Knowledge Matrix and its underlying society guidance.
How AIIE uses it
AIIE Clinical Knowledge Matrix scenario/variant ratings derived from published society guidelines.
Patient risk factors
Weight 0.10 · 10%Population risk modifiers that shift baseline pretest probability and acceptable-risk thresholds for the proposed study.
How AIIE uses it
Population risk modifiers (age, immunosuppression, pregnancy) per consensus screening guidance.
Radiation exposure burden
Weight 0.05 · 5%The ionizing-radiation cost of the proposed modality weighed against lower-energy alternatives that could answer the same question.
How AIIE uses it
BEIR / ICRP-informed trade-offs for ionizing modalities versus lower-energy alternatives.
Context inputs
Several model inputs — patient age and sex, the proposed modality and body site, the indication category, order urgency, recency of prior imaging, comorbidity burden, and the problem-list documentation signal — are model context, not standalone clinical directives. They calibrate how other factors are read; none of them alone mandates or forbids a study. The rationale for each is reproduced verbatim from the Feature Rationale Catalogue.
Patient age (years)
Age is a context input to the appropriateness model: it shifts baseline pretest probability and which indication-specific variant applies (for example, pediatric RLQ pathways versus geriatric fall protocols). Age alone does not mandate or forbid imaging; the model uses it to calibrate scores alongside indication, red flags, and proposed modality rather than as a standalone clinical recommendation.
Patient sex
Biological sex is encoded as a context modifier for the model: it affects pelvic and pregnancy-related appropriateness checks, sex-specific pretest probabilities for certain indications, and how confidently some chart-derived features are interpreted. The feature informs calibration of the score vector; it is not itself a guideline-based imaging recommendation and must be read with the full clinical scenario.
Proposed imaging modality
The ordered modality (CT, MRI, radiograph, ultrasound, and so on) is a core model input describing what study is being requested. Appropriateness depends on whether that modality can answer the clinical question with acceptable risk; this catalogue entry documents the feature as structured context for the XGBoost vector, not as a substitute for indication-specific variant selection performed at scoring time.
Proposed body site
Anatomic site codes anchor duplicate-imaging and red-flag logic to the region under evaluation. The model uses body site to align prior-study features and indication mappings; this entry records the feature as contextual input derived from the draft ServiceRequest rather than as an independent guideline directive about whether to image.
Clinical indication category
Indication category groups the presenting complaint and reason codes into a coarse clinical bucket used to select guideline mappings and baseline priors in the model. It is essential context for interpreting other features (duration, red flags, prior imaging) but does not by itself state that imaging is appropriate or inappropriate—that determination emerges from the combined feature vector and linked evidence.
Order urgency (routine / urgent / stat)
Order urgency encodes how soon imaging is requested relative to clinical stability and triage category. The model treats urgency as contextual input: the same modality may be appropriate under urgent pathways for suspected emergent pathology but inappropriate when ordered STAT for stable chronic symptoms without red flags. Appropriateness must be read with indication and acuity together.
Days since most recent relevant imaging
Elapsed time since the last study contextualizes duplicate-imaging features: the same raw prior count carries different weight when the most recent exam was yesterday versus eight weeks ago. The model uses this temporal input to calibrate how strongly recent prior imaging should decrease appropriateness, independent of any single published day threshold.
Comorbidity burden (count)
Aggregate comorbidity count summarizes chronic disease burden from problem-list and condition resources as a model context feature. It shifts baseline risk and resilience estimates in the feature vector but is not anchored to a single indication-specific variant; interpreters should combine it with indication-specific criteria rather than treating the count alone as a reason to order or defer imaging.
Imaging referenced on problem list
Whether prior imaging appears on the active problem list is a documentation-quality signal used when inferring prior-study features from unstructured and semi-structured chart data. When imaging is explicitly listed, the model weights prior-imaging counts and recency more confidently; when absent, duplicate-imaging features may be under-detected—a context input, not a clinical appropriateness rule by itself.
For the full per-feature catalogue, see the Feature Rationale Catalogue.
Primary-evidence register
Appendix D of the Technical Compendium. These are the primary frameworks and source bodies the AIIE engine draws on. Indication-specific society guidance is cited inline in the Evidence Library; the register below lists the cross-cutting methods and authorities.
| Source | Domain | How AIIE uses it |
|---|---|---|
| RAND/UCLA Appropriateness Method | rand.org | Methodological backbone for expressing appropriateness on a 1–9 panel scale and structuring indication-specific variants. |
| GRADE (Grading of Recommendations Assessment, Development and Evaluation) | gradeworkinggroup.org | Summary-of-findings framing for symptom–test concordance and for ranking the certainty of evidence behind each factor. |
| Choosing Wisely | choosingwisely.org | Overuse signals that inform the prior-imaging-redundancy factor and de-prioritize low-yield repeat studies. |
| U.S. Preventive Services Task Force (USPSTF) | uspreventiveservicestaskforce.org | Population screening recommendations feeding the patient-risk-factors factor and age/sex-conditioned pretest priors. |
| American College of Emergency Physicians (ACEP) | acep.org | Red-flag and acuity guidance for emergency presentations that warrant expedited advanced imaging. |
| PECARN (Pediatric Emergency Care Applied Research Network) | pecarn.org | Pediatric head-injury and trauma decision support that calibrates red-flag handling for children. |
| Validated clinical decision rules (Ottawa, NEXUS, Canadian C-Spine/CT Head) | theottawarules.ca | Structured rule-out criteria that shape red-flag and trauma-severity signals for extremity, cervical-spine, and head imaging. |
| BEIR / ICRP radiation guidance | icrp.org | Dose and stochastic-risk models underpinning the radiation-exposure-burden factor and modality trade-offs. |
| American Cancer Society (ACS) | cancer.org | Cancer-screening and malignancy-context guidance informing patient-risk modifiers in oncologic presentations. |