Joshua B Smith1, Matthew Shew1, Omar A Karadaghy1, Rohit Nallani1, Kevin J Sykes1, Gregory N Gan2, Jason A Brant3, Andrés M Bur1. 1. Department of Otolaryngology-Head and Neck Surgery, University of Kansas Medical Center, Kansas City, Kansas, USA. 2. Department of Radiation Oncology, University of Kansas Medical Center, Kansas City, Kansas, USA. 3. Department of Otorhinolaryngology-Head and Neck Surgery, Hospitals of the University of Pennsylvania, Philadelphia, Pennsylvania, USA.
Abstract
BACKGROUND: Machine learning (ML) algorithms may predict patients who will require salvage total laryngectomy (STL) after primary radiotherapy with or without chemotherapy for laryngeal squamous cell carcinoma (SCC). METHODS: Patients treated for T1-T3a laryngeal SCC were identified from the National Cancer Database. Multiple ML algorithms were trained to predict which patients would go on to require STL after primary nonsurgical treatment. RESULTS: A total of 16 440 cases were included. The best classification performance was achieved with a gradient boosting algorithm, which achieved accuracy of 76.0% (95% CI 74.5-77.5) and area under the curve = 0.762. The most important variables used to construct the model were distance from residence to treating facility and days from diagnosis to start of treatment. CONCLUSION: We can identify patients likely to fail primary radiotherapy with or without chemotherapy and who will go on to require STL by applying ML techniques and argue for high-quality, multidisciplinary regionalized care.
BACKGROUND: Machine learning (ML) algorithms may predict patients who will require salvage total laryngectomy (STL) after primary radiotherapy with or without chemotherapy for laryngeal squamous cell carcinoma (SCC). METHODS:Patients treated for T1-T3a laryngeal SCC were identified from the National Cancer Database. Multiple ML algorithms were trained to predict which patients would go on to require STL after primary nonsurgical treatment. RESULTS: A total of 16 440 cases were included. The best classification performance was achieved with a gradient boosting algorithm, which achieved accuracy of 76.0% (95% CI 74.5-77.5) and area under the curve = 0.762. The most important variables used to construct the model were distance from residence to treating facility and days from diagnosis to start of treatment. CONCLUSION: We can identify patients likely to fail primary radiotherapy with or without chemotherapy and who will go on to require STL by applying ML techniques and argue for high-quality, multidisciplinary regionalized care.