| Part I: Evaluation of artificial intelligence decision system for treatment of breast cancerChapter 1 Feasibility analysis of applying artificial intelligence to breast cancer treatment decisions in different subtypes and stagesPurpose: To explore the feasibility and applicability of artificial intelligence decision system in the clinical application of breast cancer patients in China.The consistency between the recommended treatment of artificial intelligence and the actual clinical treatment is initially compared,and a reasonable model of the subsequent application of artificial intelligence system in clinical is explored.Materials and Methods: The study included female invasive breast cancer patients of different stages and different categories treated by our department from June1 st to 30 th,2017.The collected clinical and pathological information of patients is imported into the WFO system and intelligent decisions are made.Early breast cancer makes retrospective decisions on adjuvant treatment,and advanced breast cancer makes retrospective decisions on first-line rescue treatment.Cases where decisions are generated successfully and those that cannot be recorded are recorded,and the concordance rate between artificial intelligence treatment and actual treatment is compared.Results are reported as rates.Results: A total of 436 breast cancer patients were screened,and 47.1%(n=205)of the cases could successfully generate artificial intelligence recommendations.Among them,149 early patients were screened,and 76 people could generate intelligent decision-making for adjuvant therapy,with an applicable rate of 51.3%.Screening 287 advanced patients,129 Watson can generate intelligent decisions for rescue treatment,the applicable rate is 45%.The reasons for the inapplicability of the analysis were mainly that the intelligent system could not identify the previous treatment plan(72.3%),the pathological type was not consistent(13.4%),and the intelligent system failed to evaluate the lesion(12.6%).The overall concordance rate between artificial intelligence recommended treatments and actual treatments was 45.8%.The concordance rate of early breast cancer is higher than that of advanced breast cancer(55.2% vs 40.3%,p=0.038).Conclusion: The application of artificial intelligence decision-making system in the clinical practice of breast cancer in China has potential application value.It is initially seen that the artificial intelligence recommended scheme and the actual scheme have a certain agreement rate.It is suggested that future research needs to be conducted for the appropriate application population,and the normativeness and value of artificialintelligence decision systems need to be further evaluated.Chapter 2 AI treatment decision support for complex breast cancer among oncologists with varying expertisePurpose: To assess treatment concordance and adherence to breast cancer treatment guidelines between oncologists and an artificial intelligence(AI)advisory tool.Materials and Methods: Study cases(n=1,977)were obtained from the Chinese Society for Clinical Oncology’s breast cancer database(2012-2017)that were at high risk for recurrence or with metastatic disease and had cell types for which the advisory tool had been trained.A cross-sectional observational study was performed to examine treatment concordance and guideline adherence among an AI advisory tool and ten oncologists with varying expertise: three fellows,four attending and three chief physicians.In a blinded fashion,each oncologist provided treatment advice on an average of 198 cases,and the advisory tool on all cases(n=1,977).Results are reported as rates and logistic regression odds ratios.Results: Among all 1977 breast cancer cases,1192 cases of early breast cancer were treated with adjuvant therapy,396 cases of advanced breast cancer were treated with first-line rescue treatment,and 389 cases of advanced breast cancer were treated with second-line rescue treatment.Concordance for recommended treatment was 0.56 for all physicians.The concordance rate was different in different stages of breast cancer.The concordance rate of first-line treatment for stage II,III,and IV breast cancer was higher than that of stage I breast cancer(0.66,0.56,0.53 compared to 0.45,p<0.001).The concordance rate is different in breast cancers with different molecular types,the lowest is in the HR / HER2-positive breast cancer(0.48),and the highest in the triple-negative breast cancer(0.69).For different treatment stages,adjuvant radiotherapy and adjuvant targeted therapy have the highest concordance rate(both greater than 0.95),adjuvant endocrine therapy and adjuvant chemotherapy have a moderate concordance rate(0.81 and 0.78,respectively),and the concordance rate of first-line rescue and second-line rescue therapy are the lowest(0.52 and 0.50,respectively).The concordance rate are higher for fellows compared to chief and attending physicians(0.68 vs.0.54,0.49;p<0.001).Adherence to treatment guidelines was significantly higher for oncologists compared to the advisory tool(0.96 vs.0.81;p<0.003),and lower for fellows compared to attending physicians(0.93 v.0.98;0.96;p=0.04).The guideline adherence rate of the AI system is affected by the stage and molecular classification.The guideline adherence rate is higher in stage I breast cancer(0.9 vs 0.57 and 0.78,p <0.001)and triple negative breast cancer(0.92 vs 0.83,p<0.001).Conclusion: The concordance rate between artificial intelligence’srecommendation and oncologist’s recommendation is affected by breast cancer staging and classification.Through the breast cancer professional guide,the normative difference between artificial intelligence decision system and doctor recommendation at different levels was verified for the first time.This study suggests that the artificial intelligence decision-making system shows good feasibility and standardization in breast cancer treatment,but there is still a certain gap between the decision-making level and high-level breast cancer specialists.Additional research in different practice settings is needed to understand the tool’s scalability and its impact on treatment decisions and clinical and health services outcomes.Chapter 3 Exploratory Research on the Impact of Artificial Intelligence on oncologist’ DecisionsPurpose: Explore the impact of artificial intelligence decision-making systems on oncologists ’treatment decisions,and evaluate the impact of artificial intelligence system-assisted doctors’ decision-making modes on treatment norms.Materials and Methods: A cross-sectional observational study was conducted using 1,977 high risk for recurrence or metastatic breast cancer cases from the Chinese Society of Clinical Oncology.Ten oncologists provided blinded treatment recommendations for breast cancer cases before and after using the artificial intelligence tool.Treatment changes are reported as rates and the effects on treatment changes of oncology experience(years),patient age and receptor subtype-TNM stage are reported as logistic regression odds ratios(ORs,95%CI).Results: Treatment decisions changed in 105 of 1,977 cases(5%)and were concentrated in hormone receptor positive(HR+)and stage IV 1st line therapy cases(73%,58%,respectively).Logistic regressions showed that decision changes were more likely in HR+ cancers(odds ratio,OR: 1.58;p<0.05),and less likely in stages IIA(OR:0.29;p<0.05)and IIIA tumors(OR: 0.08;p<0.01).Reasons for changes included the evidence provided by the tool(63% cases),patient factors highlighted by the tool(23%),and the tool’s decision logic(13%).Patient age and oncologists’ experience were not associated with decision changes.Adherence to NCCN treatment guidelines increased slightly after using the advisory tool(0.5%;p=0.003),especially in HR+HER2-breast cancer(0.77%,p<0.001).Conclusions: The artificial intelligence decision-making system has certain influence on the oncologist’s formulation of the plan,especially in the HR-positive breast cancer.With the assistance of artificial intelligence,the guideline adherence rate of doctors’ recommended treatment will be significantly improved,suggesting that artificial intelligence to assist doctors in making decisions is a currently effective application mode.Additional research on decision impact,patient-physician communication,learning and clinical outcomes are needed to establish the technology’soverall value.Part II: an exploratory study on the prediction of chemotherapy bone marrow toxicity based on artificial intelligence algorithmChapter 4 long-acting G-CSF for prophylaxis of chemotherapy-induced neutropenia in patients with breast cancer: a randomized,multicenter,active-controlled phase III trialPurpose: This study was to evaluate the efficacy and safety of long-acting G-CSF HHPG-19 K for reducing neutropenia compared with rh G-CSF.Materials and Methods:This was a randomized,controlled non-inferiority study.A total of 339 breast cancer patients who were eligible for(neo)adjuvant chemotherapy were randomized assigned into three groups to receive HHPG-19 K 100 μ g/kg,HHPG-19 K fixed dose of 6 mg or filgrastim 5 μ g/kg/day in the first cycle of chemotherapy.The primary endpoint was the duration of grade ≥ 3 neutropenia in cycle 1.The secondary endpoints included the duration of grade ≥3 neutropenia in cycles 2–4,incidence of grade ≥ 3 neutropenia,and febrile neutropenia(FN).The safety profile was also evaluated.Results: The mean duration of grade ≥3 neutropenia was 1.06 [95% confidence interval(CI): 0.65,1.26] days in HHPG-19 K 100 μg/kg group,1.23(95% CI: 0.84,1.88)days in HHPG-19 K 6 mg group,and 2.06(95% CI: 1.66,2.46)days in the filgrastim group.The mean difference between HHPG-19 K 100 μg/kg and filgrastim was –1.00(95% CI: –1.52,–0.48),the mean difference between HHPG-19 K 6 mg and filgrastim was –0.83(95% CI: –1.36,–0.30).The upper bounds of 95% CI for the difference between HHPG-19 K and filgrastim were all <1 day(the predefined non-inferiority margin).For the incidence of grade ≥3 and grade 4 neutropenia,the mean duration of grade 4 neutropenia,HHPG-19 K showed better performance compared with filgrastim.For the incidence of FN,there was no difference between patients treated with HHPG-19 K and filgrastim.For safety profile,HHPG-19 K of two doses groups were all well-tolerated.Conclusion: Long-acting HHPG-19 K is very effective and well tolerated when administered in the primary prophylaxis of chemotherapy induced neutropenia and in consecutive-cycle treatment.The fixed 6 mg-dose regimen showed similar efficacy and safety profile compared with 100 μ g/kg regimen,and would be the preference in clinical practice,due to the convenient once-per-cycle administration and high-degree treatment compliance for the patients.This study provided new evidence for the novel long-acting rh G-CSF,HHPG-19 K,which would be a new alternative for clinical practice for prophylaxis of chemotherapy induced neutropenia.Chapter 5 Exploratory research based on machine learning to predict bone marrow toxicity for breast cancer patients during chemotherapyPurpose: Based on breast cancer cases undergoing chemotherapy for the first time,to initially establish a binary classification prediction model for neutropenia in the first cycle of chemotherapy,and to evaluate the feasibility and prediction efficiency of machine learning algorithms.Materials and Methods: Based on the CSCO BC database,this study included breast cancer patients who received neoadjuvant / adjuvant chemotherapy for the first time from January 2011 to September 2015,with detailed treatment and blood routine information.The prediction models were constructed using Light GBM and other methods based on machine learning,respectively,and the predictive value of the three methods for grade 4 neutropenia was compared.Results: A total of 1083 breast cancer patients were enrolled in the study,and the data was divided into a training set and a testing set at a ratio of 8: 2.The training set included 866 cases and the testing set included 217 cases.The two groups were balanced ing age,height,weight,baseline blood routine,liver function,chemotherapy regimen,protection mode and other clinical factors,which can be used to build and verify predictive models.The accuracy rates of the three methods of Light GBM,XGBoost,and Bernoulli NB are 0.67,0.58 and 0.61,respectively.The AUC of the three methods are 0.68,0.60 and 0.62,respectively.Conclusion: The prediction model(grade 4 neutropenia)based on light GBM machine learning method for breast cancer patients after chemotherapy has higher prediction efficiency than other algorithms.In the future,this model method can be further applied to multi-class prediction of bone marrow toxicity from chemotherapy. |
PDF Full Text Request