{"id":"W2077240755","doi":"10.1016/j.annemergmed.2011.06.423","title":"391 Between-Rater Reliability of Historical Variables Used in a Routine Emergency Department Patient Evaluation","year":2011,"lang":"en","type":"article","venue":"Annals of Emergency Medicine","topic":"Reliability and Agreement in Measurement","field":"Decision Sciences","cited_by":0,"is_retracted":false,"has_abstract":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Medicine; Emergency department; Inter-rater reliability; Reliability (semiconductor); Medical emergency; Emergency medicine; Statistics; Nursing","routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false,"invisible_to_affiliation_only":false},"retraction":null,"screen":null,"direct_labels":[],"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.01992224,0.0002649496,0.0008803452,0.0003949937,0.00005854763,0.000001682033,0.0006562144,0.0001274916,0.02057719],"category_scores_gemma":[0.01046683,0.0001761178,0.0002481335,0.001369104,0.0001066689,0.0002424662,0.0001223979,0.0001659072,0.00002553469],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001361042,"about_ca_system_score_gemma":0.0001021719,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.001582397,"about_ca_topic_score_gemma":0.0001937873,"domain_scores_codex":[0.9894651,0.001100993,0.004259087,0.000703656,0.00406644,0.0004047196],"domain_scores_gemma":[0.9946095,0.0003328653,0.001205116,0.001128156,0.002538831,0.0001855306],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0001540185,0.001052499,0.9504666,0.00006135655,0.00004463052,8.689103e-7,0.003809755,0.0001250818,0.001334457,0.0002288327,0.03738501,0.005336937],"study_design_scores_gemma":[0.0009700782,0.002327125,0.8937091,0.0001550933,0.000128513,1.846854e-7,0.0006770993,0.0003795357,0.008904382,0.08381177,0.008643134,0.0002939391],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9874308,0.001211313,0.0006952647,0.002194338,0.002513865,0.0008814094,0.00001830698,0.00001187763,0.005042864],"genre_scores_gemma":[0.9987464,0.0004747722,0.0002806434,0.00003156525,0.0001341206,0.00006626291,0.00001728126,0.00001011564,0.0002388447],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.08358294,"threshold_uncertainty_score":0.9978684,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.5818926185803324,"score_gpt":0.4632303106880315,"score_spread":0.1186623078923009,"validation_status":"score_only:v0-immature-baseline","note":"Baseline scores from an immature model (maturity gate not passed). Scores rank; they never assert a category."}}