{"id":"W3199377622","doi":"10.1002/pmrj.12705","title":"Reliability and validity of subjective radiologist reporting of temporal changes in lumbar spine <scp>MRI</scp> findings","year":2021,"lang":"en","type":"article","venue":"PM&R","topic":"Radiology practices and education","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"ca_institutions":"Western University","funders":"National Institute of Arthritis and Musculoskeletal and Skin Diseases; National Health and Medical Research Council; National Institutes of Health","keywords":"Medicine; Magnetic resonance imaging; Inter-rater reliability; Lumbar; Low back pain; Radiology; Kappa; Reliability (semiconductor); False positive paradox; Physical therapy; Physical medicine and rehabilitation; Psychology; Pathology; Rating scale","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":"metacan-v3-hybrid-931329e0061c","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.04332433,0.0004163071,0.0006001533,0.001565326,0.0004655076,0.001399584,0.000799197,0.0005702616,0.0008361154],"category_scores_gemma":[0.1150494,0.0004415654,0.0008368571,0.000558171,0.001466437,0.001094479,0.001995999,0.0004391441,0.0002655858],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0003796631,"about_ca_system_score_gemma":0.0004728233,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.001539856,"about_ca_topic_score_gemma":0.002031389,"domain_scores_codex":[0.9648178,0.02024748,0.004538221,0.002916787,0.006815084,0.0006645497],"domain_scores_gemma":[0.8413358,0.0894243,0.02796535,0.0112794,0.02854177,0.001453276],"domain_codex":null,"domain_gemma":"reporting","domain_candidate":"reporting","domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.001192257,0.00009782406,0.9716636,0.0004485922,0.0007934567,0.0001227916,0.006116667,0.0005628168,0.002124967,0.0001314855,0.0003479199,0.01639747],"study_design_scores_gemma":[0.00007497283,0.0008746298,0.9866194,0.0002757393,0.000296198,0.0007712552,0.003885709,0.00388734,0.002090626,0.0003274194,0.0008162506,0.00008047324],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9931792,0.000552347,0.004320415,0.00005171212,0.00005983342,0.0001296121,0.0001826006,0.00002905164,0.001495364],"genre_scores_gemma":[0.9977173,0.0001024403,0.001775352,0.00002723453,0.00002861016,0.00009316434,0.0001397019,0.00001111682,0.0001051396],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.9566756,"threshold_uncertainty_score":0.2291237,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.05999772335550631,"score_gpt":0.3460465078143599,"score_spread":0.2860487844588536,"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."}}