1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 Methods Matter: Clinical Prediction Models Will Benefit Sports Medicine Practice, But Only if They Are Properly Developed and Validated *Garrett S. Bullock, PT, DPT,1,2 *Tom Hughes, PT, PhD3,4 Jamie C. Sergeant, MSci, DPhil5,6 Michael J. Callaghan, MCSP, MPhil, DPhil3,4,5 Richard D. Riley, PhD7 Gary S. Collins, PhD8,9 1. Centre for Sport, Exercise and Osteoarthritis Research Versus Arthritis, University of Oxford, United Kingdom 2. Nuffield Department of Orthopaedics, Rheumatology, and Musculoskeletal Sciences, University of Oxford, Oxford, United Kingdom 3. Manchester United Football Club, AON Training Complex, Birch Road, Off Isherwood Road, Carrington, Manchester M31 4BH, UK 4. Department of Health Professions, Manchester Metropolitan University, Manchester, UK. 5. Centre for Epidemiology Versus Arthritis, Centre for Musculoskeletal Research, Manchester Academic Health Science Centre, University of Manchester, Manchester, UK 6. Centre for Biostatistics, Manchester Academic Health Science Centre, University of Manchester, Manchester, UK. 7. Centre for Prognosis Research, School of Medicine, Keele University, Staffordshire, UK. 8. Centre for Statistics in Medicine, Nuffield Department of Orthopaedics, Rheumatology, and Musculoskeletal Sciences, University of Oxford, Oxford UK 9. Oxford University Hospitals NHS Foundation Trust, Oxford, United Kingdom *Joint first authors Corresponding author: Garrett S. Bullock PT, DPT Nuffield Department of Orthopaedics, Rheumatology, and Musculoskeletal Sciences University of Oxford Oxford, United Kingdom OX3 7LD (865) 227-374 [email protected] 1 35 Introduction 36 Sports medicine clinicians are expected to make accurate diagnoses, estimate prognoses, and 37 identify athletes at risk of sustaining an injury 1. These complex decisions are dependent on 38 clinical reasoning, which is informed by, and often biased toward, a practitioner’s scientific 39 knowledge and experience. Clinical prediction models are developed by researchers to help 40 facilitate such decisions in practice 2; data for multiple predictor variables are combined to 41 estimate an individual’s risk of a health outcome either being present (diagnosis) or whether it 42 will occur in future (prognosis) 3. Despite being employed widely in clinical medicine, clinical 43 prediction models are uncommon in sports medicine. Clinical prediction models can offer 44 benefits to both practitioners and athletes, but only if they are developed and validated using 45 rigorous methods and transparently reported so that potential users can judge their accuracy and 46 usefulness. 47 48 Therefore, the purpose of this editorial is to describe the recommended steps for clinical 49 prediction development and validation and to guide practitioners using and interpreting 50 prediction models in sports medicine. 51 52 53 Model Development 54 The first step in developing a prediction model is to identify its clinical need, the target 55 population, and how and when it would fit into the clinical workflow. Models should predict 56 outcomes that are relevant to sport stakeholders, and be clearly defined, including how and when 57 assessed 4. 2 58 59 Next is to identify any existing models that could be evaluated or updated. If not, then before 60 developing a new model, a publicly accessible protocol should be developed 4. A summary of 61 the recommended steps is in Table 1 3. 62 63 The natural design for developing a prediction model is a cross-sectional study for developing a 64 diagnostic model and a longitudinal study for a prognostic model 3. For the latter, follow-up 65 periods should be of sufficient duration to measure the desired outcome. Datasets used to 66 develop prediction models are rarely complete. Omitting individuals with incomplete data should 67 be avoided, as it reduces the sample size and may lead to bias. Multiple imputation should be 68 considered for handling missing data 2,3. 69 70 Typically, many predictors are available for potential inclusion in a prediction model and 71 reduction is often needed. Omitting predictors based on univariable association with the outcome 72 should be avoided. Instead, predictors considered for inclusion should be identified based upon 73 existing evidence, and clinical reasoning to determine their importance, re