Marlon Perera1, Lewis Smith2, Ian Thompson3, Geoff Breemer2, Nathan Papa4, Manish I Patel5, Peter Swindle6, Elliot Smith2. 1. Department of Urology, Mater Hospital, Brisbane, Australia; Department of Urology, Memorial Sloan Kettering Cancer Center, New York, NY, USA; Faculty of Medicine, University of Queensland, Brisbane, Australia. Electronic address: marlonlperera@gmail.com. 2. Maxwell Plus, Brisbane, Australia. 3. Department of Urology, Christus Health, San Antonio, TX, USA. 4. Department of Public Health and Preventative Medicine, Monash University, Melbourne, Australia. 5. Department of Urology, Westmead Hospital, Faculty of Medicine and Health, The University of Sydney, Sydney, Australia. 6. Department of Urology, Mater Hospital, Brisbane, Australia; Maxwell Plus, Brisbane, Australia.
Abstract
BACKGROUND: Prostate-specific antigen (PSA) kinetics, defined as the change in PSA over time, may be of use as a predictor of prostate cancer. PSA kinetics can be assessed as the PSA velocity, which is traditionally evaluated dichotomously and classified as abnormal if greater than either 0.35 or 0.75 ng/ml/yr. Machine learning models may provide additional benefit in assessing risk using PSA kinetics instead of PSA velocity. OBJECTIVE: To improve the utility of PSA kinetics by constructing a generalizable, universal machine learning model. DESIGN, SETTING, AND PARTICIPANTS: Data were obtained from the PLCO and PCPT trials and from a contemporary Australian cohort. PSA data were interpolated using a modified Gaussian process. A machine learning model based on a two-headed approach was designed, in which the multivariable input was fed into a one-dimensional ResNet18 model. OUTCOME MEASURES AND STATISTICAL ANALYSIS: The model performance was assessed compared to PSA levels and PSA velocity in terms of area under the receiver operator characteristic curve (AUC). RESULTS AND LIMITATIONS: A total of 10719 patients were included in the analysis. In tests on a validation set of the complete database to diagnose grade group ≥2, the AUC was 0.886 (95% confidence interval [CI] 0.870-0.902) for the machine learning model, compared to 0.807 (95% CI 0.796-0.819) for PSA and 0.627 (95% CI 0.607-0.648) for PSA velocity. CONCLUSIONS: Machine learning models can be used to augment the diagnostic utility of PSA kinetics in the diagnosis of prostate cancer. We demonstrated significant improvements in accuracy compared to the traditional approaches of PSA velocity and PSA thresholds. PATIENT SUMMARY: Prostate cancer diagnosis is limited by the diagnostic accuracy of the prostate-specific antigen (PSA) blood test. Advances in techniques such as machine learning algorithms can greatly improve the diagnostic accuracy of prostate cancer screening without additional costs or tests.
BACKGROUND: Prostate-specific antigen (PSA) kinetics, defined as the change in PSA over time, may be of use as a predictor of prostate cancer. PSA kinetics can be assessed as the PSA velocity, which is traditionally evaluated dichotomously and classified as abnormal if greater than either 0.35 or 0.75 ng/ml/yr. Machine learning models may provide additional benefit in assessing risk using PSA kinetics instead of PSA velocity. OBJECTIVE: To improve the utility of PSA kinetics by constructing a generalizable, universal machine learning model. DESIGN, SETTING, AND PARTICIPANTS: Data were obtained from the PLCO and PCPT trials and from a contemporary Australian cohort. PSA data were interpolated using a modified Gaussian process. A machine learning model based on a two-headed approach was designed, in which the multivariable input was fed into a one-dimensional ResNet18 model. OUTCOME MEASURES AND STATISTICAL ANALYSIS: The model performance was assessed compared to PSA levels and PSA velocity in terms of area under the receiver operator characteristic curve (AUC). RESULTS AND LIMITATIONS: A total of 10719 patients were included in the analysis. In tests on a validation set of the complete database to diagnose grade group ≥2, the AUC was 0.886 (95% confidence interval [CI] 0.870-0.902) for the machine learning model, compared to 0.807 (95% CI 0.796-0.819) for PSA and 0.627 (95% CI 0.607-0.648) for PSA velocity. CONCLUSIONS: Machine learning models can be used to augment the diagnostic utility of PSA kinetics in the diagnosis of prostate cancer. We demonstrated significant improvements in accuracy compared to the traditional approaches of PSA velocity and PSA thresholds. PATIENT SUMMARY: Prostate cancer diagnosis is limited by the diagnostic accuracy of the prostate-specific antigen (PSA) blood test. Advances in techniques such as machine learning algorithms can greatly improve the diagnostic accuracy of prostate cancer screening without additional costs or tests.
Authors: Ian M Thompson; Donna Pauler Ankerst; Chen Chi; Phyllis J Goodman; Catherine M Tangen; M Scott Lucia; Ziding Feng; Howard L Parnes; Charles A Coltman Journal: J Natl Cancer Inst Date: 2006-04-19 Impact factor: 13.506
Authors: Andrew J Vickers; Cathee Till; Catherine M Tangen; Hans Lilja; Ian M Thompson Journal: J Natl Cancer Inst Date: 2011-02-24 Impact factor: 13.506
Authors: H B Carter; C H Morrell; J D Pearson; L J Brant; C C Plato; E J Metter; D W Chan; J L Fozard; P C Walsh Journal: Cancer Res Date: 1992-06-15 Impact factor: 12.701
Authors: Fritz H Schröder; Jonas Hugosson; Monique J Roobol; Teuvo L J Tammela; Marco Zappa; Vera Nelen; Maciej Kwiatkowski; Marcos Lujan; Liisa Määttänen; Hans Lilja; Louis J Denis; Franz Recker; Alvaro Paez; Chris H Bangma; Sigrid Carlsson; Donella Puliti; Arnauld Villers; Xavier Rebillard; Matti Hakama; Ulf-Hakan Stenman; Paula Kujala; Kimmo Taari; Gunnar Aus; Andreas Huber; Theo H van der Kwast; Ron H N van Schaik; Harry J de Koning; Sue M Moss; Anssi Auvinen Journal: Lancet Date: 2014-08-06 Impact factor: 79.321