Page 5 - CUA2019 Abstracts - Oncology-Kidney

2019 CUA Abstracts

MP-9.13 surgical candidates or due to patient preference. This study describes
Machine learning to predict recurrence of localized renal cell functional and oncological outcomes of ablation therapy for SRMs at
carcinoma our centre.
1
1
Yanbo Guo , Luis H. Braga , Anil Kapoor 1 Methods: A total of 166 patients who underwent ablation therapy for SRM
1 Urology, McMaster University, Hamilton, ON, Canada at London Health Sciences Centre (LHSC) between 2011 and 2017 were
Introduction: The increased incidence of renal cell carcinoma (RCC) has retrospectively reviewed. Ablation therapy included radiofrequency abla-
been largely explained by the increased use of diagnostic imaging and tion (RFA), cryoablation (CRA), and microwave ablation (MWA). Patients
discovery of localized disease. Localized RCC has a five-year survival rate with renal lesions ≤4 cm with recorded followup (FU) to 12 months were
1
of nearly 90% but there remains a 20–30% risk of recurrence. Current included. Patients with simultaneous multiple renal lesions, known meta-
guidelines stratify patients between risk categories based upon their static diseases, or ablations for recurrences were excluded. Oncological
pathologic grade and TMN stage. 2,3 Nomograms that incorporate other and functional outcomes were assessed.
variables are available but they also rely mainly upon pathologic findings. Results: Median FU was 25 months (interquartile range [IQR] 13–41). Most
Our objective is to use a cloud-based machine learning (ML) platform to patients (70%) were male. Mean age was 68.2 years (standard deviation
develop a model for recurrence after curative treatment of localized RCC [SD] 10.6) with a mean body mass index (BMI) of 30.7 (SD 7.9); 8.5%
using pre- and postoperative variables. had solitary kidney. Median Charlson comorbidity index was 5 (IQR 4–6).
Methods: A de-identified RCC database from our institution was uploaded Mean tumour diameter was 2.6 cm (SD 0.8). A total of 62.9%, 33.1%,
®
to the Microsoft Azure Machine Learning Studio. The dataset was then and 4.0% of patients had low, intermediate, and high RENAL nephrometry
split into a training and a testing group. Two ML models were trained, a scores, respectively. RFA occurred in 112 patients, 47 patients underwent
two-class neuro network model and a two-class boosted decision tree CRA, and seven underwent MWA. Biopsy showed clear-cell histology
model, both fundamental approaches in ML. These models were then (63.4 %), papillary (21.7%), chromophobe (6.9%), and oncocytoma (6.1%).
4
evaluated using the area under curve (AUC) of a ROC curve. There was no difference in serum creatinine post- and pre-ablation (112.3
Results: A total of 697 patients were part of the dataset. Seven variables vs. 100.2; p=0.13). Complete radiographic response was seen in 81.9%
were included in our model. The optimized model achieved an AUC of and 10% needed repeated ablation for residual disease; 11.3% had local
0.877. Setting a threshold to maximize sensitivity, there was a sensitivity recurrences (RFA=8 vs. CRA=8 vs. MWA=2; p=0.089). Two patients died
of 89.47%, a specificity of 71.95 %, and positive predictive value of 3.19. of progression and metastasis. Six patients had Clavien 1, three patients
Conclusions: We built an accurate RCC recurrence prediction model had Clavien II, and four patients had Clavien III complications (one urine
using an accessible cloud-based ML platform. This approach offers advan- leak, two ureteral injuries, and one pneumothorax).
tages over traditional statistics, including the ability to easily incorpo- Conclusions: Ablation therapy, with different available modalities at our
rate new data and distribute updates. Our dataset is a part of a larger institution, is a viable option with a low-risk profile and low recurrence rates.
national dataset, which we aim to incorporate into future iterations.
Currently, this model’s performance favourably compares to existing UP-9.2
nomograms. With more accurate prognostication of recurrence, we can Clinical features and outcomes of secondary somatic malignancy
5
better counsel patients, individualize surveillance strategies, minimize arising from teratoma in late relapse germ cell tumour
ineffective investigations, and identify high-risk patients who truly benefit Nathan C. Wong , Shawn R. Dason , Lucas Dean , Sumit Isharwal , Mark
2
1
1
1
from close followup. Donoghue , Liwei Jia , William Tap , Gabriella Joseph , Samuel Funt ,
5
5
3
5
4
References Deaglan McHugh , Hikmat Al-Ahmadie , Victor E. Reuter , Robert J.
5
4
4
1. Chin AI, Lam JS, Figlin RA, et al. Surveillance strategies for renal cell Motzer , George J. Bosl , Joel Sheinfeld , David B. Solit , Darren R. Feldman 5
1
5
5
5
carcinoma patients following nephrectomy. Rev Urol 2006;8:1-7. 1 Department of Surgery, Urology Services, Memorial Sloan Kettering Cancer
2. Kassouf W, Monteiro LL, Drachenberg DE, et al. Canadian Urological Center, New York, NY, United States; Department of Urology, University of
2
Association guidelines for followup of patients after treatment of Virginia Health System, Charlottesville, VA, United States; Department of
3
non-metastatic renal cell carcinoma. Can Urol Assoc J 2018;12:231- Molecular Oncology, Memorial Sloan Kettering Cancer Center, New York,
8. https://doi.org/10.5489/cuaj.5462 NY, United States; Department of Pathology, Memorial Sloan Kettering
4
3. Ljungberg B, Bensalah K, Canfield S, et al. EAU guidelines on renal Cancer Center, New York, NY, United States; Department of Medicine,
5
cell carcinoma: 2014 update. Eur Urol 2015;67:913-24. https://doi. Memorial Sloan Kettering Cancer Center, New York, NY, United States
org/10.1016/j.eururo.2015.01.005 Introduction: Late relapse (LR) (>2 years) germ cell tumour (GCT) is
4. Kotsiantis SB. Supervised machine learning: A review of classifica- associated with an increased rate of secondary somatic malignancy (SSM).
tion techniques. Informatica 2007;31. We report our experience with SSM in the setting of LR and determine
5. Sorbellini M, Kattan MW, Snyder ME, et al. A postoperative prognos- predictors of overall survival (OS).
tic nomogram predicting recurrence for patients with conventional Methods: From 1985–2018, 46 patients with GCT and SSM at LR were
clear cell renal cell carcinoma. J Urol 2005;173:48-51. https://doi. identified and underwent chart review. The Kaplan-Meier method was
org/10.1097/01.ju.0000148261.19532.2c used to estimate OS from time of LR and a Cox proportional hazards
6. Sorbellini M, Kattan MW, Snyder ME, et al. A postoperative prognos- model to assess predictors of OS.
tic nomogram predicting recurrence for patients with conventional Results: Of 46 men (44 testicular primary, two mediastinal), median time
clear-cell renal cell carcinoma. J Urol 2005;173:48-51. https://doi. to LR with SSM was 10.4 years. Most (n=27) were symptomatic at pre-
org/10.1097/01.ju.0000148261.19532.2c sentation but 11 were detected by elevated tumour markers and eight by
imaging. SSMs were predominately adenocarcinoma (25), sarcoma (14),
UP-9.1 Wilms tumour (two), primitive neuroectodermal tumour (PNET) (one),
Outcomes of ablation therapy for small renal masses from a and other (four). Median time to LR was longer for adenocarcinoma vs.
single centre other histotypes of SSM (14.6 vs. 4.1 years; p<0.001). The initial site of
1
1
Samir Sami , Shiva Nair , Lucy Saimalov , Khalil Hetou , Stephen E. LR was the retroperitoneum (RP, 26), pelvis (seven), lung (six), retrocrural
2
1
Pautler , Amol Mujoomdar , Joseph Chin 1 space (three), mediastinum (two), and other (two). Only 10/26 men with
3
1
1 Division of Urology, Western University, London, ON, Canada; LR in the RP had undergone prior retroperitoneal lymph node dissection
2 Medicine, Western University, London, ON, Canada; Radiology, Western (RPLND) (all at outside institutions; variable templates) with teratoma in
3
University, London, ON, Canada 7/10. The other 16 men had received chemotherapy only (eight), orchiec-
Introduction: The prevalence of imaging has lead to an increase in the tomy only for stage I (three), RPLND aborted (one), and unknown (four).
incidence of small renal masses (SRMs). SRMs presumed to be malignant All LR were managed with surgery; 26 also received chemotherapy (16
are most commonly treated with a partial nephrectomy. Non-invasive SSM-directed, 10 GCT-directed). Overall, 12 patients died and the median
ablative techniques are increasingly used in patients (pts) who are poor OS was 14.2 years. On univariable analysis, symptomatic presentation
S144 CUAJ • June 2019 • Volume 13, Issue 6(Suppl5)