Analytics environment
Three planned modes of analysis on the HARK release.
Natural-language queries that compile to auditable SQL; graph analyses over the harmonized clinical record; and hypothesis-generating association mining. The first is demonstrated below on synthetic data. The latter two are planned capabilities of the consortium's analytics layer.
Capability · 01
Natural-language query
The first analytics layer translates English questions into auditable SQL, runs them against the HARK release, and returns a chart, the generated SQL, and a written finding. The demonstration below runs on a synthetic dataset; nothing on this page reflects real patient records.
Try an example
Coming soon
Free-text mode and bring-your-own-API-key are next on the roadmap. For now the demo runs precomputed responses so the public site stays free.
Capability · 02 · Planned capability
Graph analyses over the harmonized record
Audiometric care generates a network: patients linked to encounters, encounters linked to measurements, measurements linked to ontology concepts, and concepts linked through external reference graphs (LOINC, SNOMED CT, OMOP). Hosting HARK on shared infrastructure enables analyses that traverse this network rather than collapsing it into a single table.
Comorbidity networks
Project the patient-by-diagnosis matrix into a concept co-occurrence graph and inspect which non-otologic conditions cluster with sensorineural hearing-loss patterns across the consortium.
“Which SNOMED CT concepts are over-represented in the 1-hop neighborhood of bilateral high-frequency sensorineural hearing loss in patients aged 50 to 65?”
Trajectory similarity
Embed patient audiometric trajectories and compute nearest-neighbor sets to find cohorts that progress similarly. Useful for normative referencing, clustering, and case-finding.
“Find the 1,000 patients whose 5-year audiometric trajectory most closely resembles this one. Stratify by audiogram configuration.”
Care-pathway graphs
Treat encounter sequences as edges in a temporal graph. Surface common pathways from first audiology evaluation to amplification, surgical referral, or discharge.
“What is the modal sequence of encounter types in the year following an asymmetric SNHL finding, and how does it differ across contributing sites?”
Concept-graph crosswalks
Walk between HARK concepts and the external ontology graphs they map to. Enables propagation of definitions, equivalence checks, and federated queries against OMOP-mapped resources.
“List every HARK measurement concept that has no LOINC equivalent and trace which SNOMED CT concepts are used as bridge mappings.”
Capability · 03 · Planned capability
Hypothesis generation and novel-association mining
Routinely collected audiometric data, when harmonized at consortium scale, are large enough to support targeted hypothesis generation. The analytics layer is being designed to support associational and predictive analyses that flag candidate relationships for follow-up study, with explicit caveats about confounding, selection, and the observational nature of the data.
Association screening
Pre-registered, multiple-testing-corrected screens for relationships between audiometric phenotypes and demographic, clinical, exposure, or treatment variables. Output is ranked candidate associations with effect sizes, intervals, and study-design caveats.
Subgroup discovery
Algorithmic search for patient subgroups whose audiometric or longitudinal behavior departs from the cohort baseline. Surfaces previously underdescribed phenotypes for confirmatory analysis.
Predictive modeling
Reference benchmarks for predicting hearing-loss progression, treatment response, and care utilization from harmonized features. Models are released alongside the data version they were trained on, with held-out evaluation across sites.
Hypothesis registry
Candidate associations and the queries that generated them are versioned in a public registry, so claims can be re-run on later HARK releases and across contributing sites.
These capabilities are framed as hypothesis generation, not causal inference. The consortium analytics environment will surface candidate signals that warrant prospective study, replication on independent data, and methodological scrutiny before clinical interpretation.
Case study · preliminary data
Predicting vestibular schwannoma from routinely collected audiometric features
Single-institution proof of concept · Mass Eye and Ear · 20,346-patient hold-out (450 VS cases, ~2.2% prevalence)
Vestibular schwannoma (VS) is a benign tumor of the vestibulocochlear nerve. It is diagnostically uncommon but high-consequence, and its audiometric signature is heterogeneous, ranging from frank asymmetric sensorineural loss to subtle thresholds and configurations with otherwise unremarkable audiometry. Whether routinely collected audiometric data alone carry enough signal to flag patients who warrant MRI follow-up is an open question.
Finding
On the held-out test set, the top-performing models reached ROC-AUC values of approximately 0.85 (voting ensemble 0.850, baseline gradient-boosted tree 0.848, anomaly-guided ensemble 0.845). Platt scaling reduced the baseline Brier score from 0.090 to 0.020. At the F1-optimal operating point (threshold 0.77), sensitivity was 0.44 and specificity 0.96 against a 2.2% case prevalence; at the Youden-optimal operating point (threshold 0.33), sensitivity rose to 0.76 with specificity 0.81. At a fixed 5% false-positive-rate ceiling, recall was 46.2%, and the top 1% of risk scores contained 25.1% true VS cases against a base rate of 2.2%. Decision-curve analysis on the matched cohort showed positive net benefit over treat-all and treat-none baselines across low, prevalence-aware thresholds (approximately 0.01 to 0.07).
What consortium scale adds
- Single-site case counts limit subgroup analysis (age, audiogram configuration, asymmetry pattern). Consortium-scale harmonization would enable site-stratified evaluation and external validation.
- Label noise is bounded by one institution's MRI-reporting conventions. Aligning labeling rules across HARK contributors would expose and correct site-specific failure modes.
- Operating thresholds chosen on a single prevalence may not generalize. A multi-site release supports prevalence-aware threshold selection and prospective performance monitoring.
- The same harmonized features support broader otologic and hearing-loss prediction tasks beyond VS, so the case study is a template for downstream consortium-scale studies.
Design notes
- All three capabilities run against the same versioned HARK release. Queries, graphs, and models can be reproduced against the exact data state they were generated on.
- Auditable SQL is generated for every natural-language question. The analytics layer assists; it does not replace investigator judgment.
- Outputs include the data version, the contributing sites included, and a written caveat block describing known limitations of the harmonized record at that release.
Continue