Publication: Erken evre meme kanseri hastaların sağkalım sürelerinin ve olasılıklarının, geleneksel sağkalım yöntemleri ile makine ögrenmesi tekniklerinin R paket programında karşılaştırlması
Abstract
Amaç: Erken evre meme kanseri hastaların klinik başvuru ve patolojik bilgilerinegöre, tedavi sonrası sağkalım olasılıklarına ilişkin R yazılımlı geleneksel sağkalımyöntemleri ile makine ögrenmesi tekniklerinin prognostik faktörler açısındantahminlenmesi ve tahmin yöntemlerinin karşılaştırılmasıdır.Gereç ve Yöntem: Çalışmaya Marmara Üniversitesi Tıbbi Onkoloji polikliniğindeerken evre meme kanseri tanısı konulan 697 hasta dâhil edilmiştir. Çalışmanın ilkaşamasında erken evre meme kanseri hastaların 5 ve 10 yıllık sağkalım olasılıklarıKaplan-Meier yöntemi ile hesaplanmıştır. Sağkalıma etkili olan anlamlı bağımsızdeğişkenlerin belirlenmesinde; tek değişkenli analizlerde anlamlı risk (hazard)faktörlerin tespiti için Log-Rank testi, çok değişkenli analizlerde ise Cox orantısalhazard (COH) regresyon çözümlenmesi yapılmıştır. Çalışmanın ikinci aşamasında ise,aynı veri setine makine öğrenmesi algortimaları olan Rastgele Orman, Xgboost veGbm modelleri uygulanarak, 5 ve 10 yıllık sağkalım olasılıkları hesaplanmış olupönem derecesine göre sağkalıma etkili olan risk (hazard) faktörler sıralanmıştır.Bulgular: Evre meme kanseri tanısı konulan 697 hastanın makine öğrenmesialgoritmalarıyla ve geleneksel sağkalım yöntemleriyle hesaplanan 5 ve 10 yıllıksağkalım olasılıkları hemen hemen aynı değerlere sahiptir. Risk (hazard) faktörleriaçısından yapılan karşılaştırmasında ise meme kanseri hastaları için önemli bir riskgöstergesi olan Her-2 faktörü makine öğrenme algoritmalarıyla oluşturulan sağkalımmodellerinin ikisinde (Gbm, Xgboost) önemli ilk 5 risk faktörün içerisinde sıralanmışolasına rağmen geleneksel yöntemlerde anlamlı çıkmamıştır.Sonuçlar: Sağlık alanına ait, özellikle takibe dayalı çalışmalarda eksik gözlem veyadengesiz dağılım sergileyen veri setlerinde geleneksel yöntemlerin yerine makineöğrenmesi algoritmalarının tercih edilmesi özellikle Gbm ile Xgboost algoritmasıanaliz sonuçlarını daha güvenilir yapacaktır.
Objective: The aim of this study is to estimate and compare the prognostic factors oftraditional survival methods with R-software and machine learning techniques basedon clinical presentation and pathological information of early stage breast cancerpatients.Materials and Methods: 697 patients diagnosed with early stage breast cancer inMarmara Oncology Outpatient Clinic of Marmara University were included in thestudy. In the first stage of the study, the 5-year and 10-year survival rates of earlystagebreast cancer patients were calculated using the Kaplan-Meier method. Indetermining significant independent variables that are effective on survival; Log-Ranktest was used for determination of significant risk (hazard) factors in univariateanalyzes, and Cox proportional hazard (COH) regression analysis was used inmultivariate analyzes. In the second stage of the study, 5 and 10-year survivalprobabilities were calculated by applying Random Forest, Xgboost and Gbm models,which are machine learning algorithms to the same data set, and the risk factors thataffect survival are listed according to their significance level.Results: The 5 and 10-year survival probabilities of 697 patients diagnosed with stagebreast cancer were approximately the same as those calculated by machine learningalgorithms and traditional survival methods. In terms of risk factors, Her-2 factor,which is an important risk indicator for breast cancer patients, was not significant intraditional methods although it was listed among the top 5 risk factors in two of thesurvival models (Gbm, Xgboost) created by machine learning algorithms.Conclusions: The use of machine learning algorithms(gbm,xgboost) instead oftraditional methods in health care, especially in follow-up studies showing incompleteobservation or unbalanced distribution will make the analysis results more reliable.
Objective: The aim of this study is to estimate and compare the prognostic factors oftraditional survival methods with R-software and machine learning techniques basedon clinical presentation and pathological information of early stage breast cancerpatients.Materials and Methods: 697 patients diagnosed with early stage breast cancer inMarmara Oncology Outpatient Clinic of Marmara University were included in thestudy. In the first stage of the study, the 5-year and 10-year survival rates of earlystagebreast cancer patients were calculated using the Kaplan-Meier method. Indetermining significant independent variables that are effective on survival; Log-Ranktest was used for determination of significant risk (hazard) factors in univariateanalyzes, and Cox proportional hazard (COH) regression analysis was used inmultivariate analyzes. In the second stage of the study, 5 and 10-year survivalprobabilities were calculated by applying Random Forest, Xgboost and Gbm models,which are machine learning algorithms to the same data set, and the risk factors thataffect survival are listed according to their significance level.Results: The 5 and 10-year survival probabilities of 697 patients diagnosed with stagebreast cancer were approximately the same as those calculated by machine learningalgorithms and traditional survival methods. In terms of risk factors, Her-2 factor,which is an important risk indicator for breast cancer patients, was not significant intraditional methods although it was listed among the top 5 risk factors in two of thesurvival models (Gbm, Xgboost) created by machine learning algorithms.Conclusions: The use of machine learning algorithms(gbm,xgboost) instead oftraditional methods in health care, especially in follow-up studies showing incompleteobservation or unbalanced distribution will make the analysis results more reliable.