Neoadjuvant chemotherapy
Patients with esophageal and GOJ type 1/2 tumours21 received either cisplatin and 5-fluorouracil (5-FU; CF; 2 cycles),22 oxaliplatin and 5-FU (OF; 2 cycles),23 epirubicin, cisplatin, and 5-FU (ECF; 3 cycles), epirubicin, cisplatin, and capecitabine (ECX; 3 or 4 cycles),24 or epirubicin, oxaliplatin, and capecitabine (EOX; 3 cycles). Patients with type 3 tumours received ECX/EOX/ECF (3 cycles). Some patients (distal oesophageal/GOJ) received 3 cycles of ECX pre-and post-operatively+/-bevacizumab,25 or 3 cycles of ECF pre- and post-operatively.26
Decision theory and cost analysis
PET-CT sensitivity for metastatic disease in this study (overt metastases plus unsuspected unresectable disease 39.5%); test risk, the additional lifetime attributable cancer risk for the “average” patient in this series adjusted for gender and age (60 years), for a 13mSV radiation dose (0.067%),38 adjusted for the likelihood of 2-year survival10 (40.0%; =0.033%); net treatment benefit of metastases (the avoidance of futile surgery; assumed by necessity to be 100%); treatment risk: inappropriate palliative management (similarly 100%); and false positive rate (assumed by necessity to be 0.00%, as PET-CT +/- confirmatory testing for equivocal lesions represents present gold-standard, and there are a lack of relevant data). For LR models for which a test Pt was not applicable (predicting unresectable disease), the optimal classification threshold was determined using pROC39.
A 15mSV radiation dose for CT was used to calculate an estimated age (60 years) and gender-adjusted lifetime cancer risk of 0.077%.12Cost analysis was performed using 2013-2014 National Health Service tariffs: esophagectomy (£12,274), laparoscopy (£1,535), PET-CT (£950), and CT (£140)16.
Model development, tuning, validation, and performance
Three techniques were used: logistic regression (LR; backwards stepwise binary logistic), decision tree analysis (DTA; recursive partitioning using loss matrices) and artificial neural networks (ANN; feed forward back-propagation multilayer perceptron).34,35 Models were tuned, generated, and validated internally (bootstrapping) as described previously and in the supplementary methods.36To avoid scanner bias, models were developed using patients’ staged/re-staged using the more recent scanner, and validated independently using patients’ staged/restaged using the earlier scanner. For models predicting progression to metastases, there were n=198 in the development group, n=185 in the validation group. For those predicting unresectable disease there were 168 and 72, respectively.
Recursive and partitioning trees partition data use variables and misclassification costs to generate intuitive decision trees and identify interactions and risk subgroups. However, they can over-fit on spurious/arbitrary associations, and ignore clinically important, but less statistically significant partitions.34 LR, whilst less able to partition, is preferable for linear associations. By contrast, ANNs identify more subtle and complex non-linear interactions, but at high risk of over-fitting.35