A retrospective review of the charts of 91 patients was performed, with prospectively collected urine samples from an existing colorectal cancer tissue bank. Adult patients with clinical stage III or IV colorectal cancer (based on preoperative imaging, colonoscopies and biopsy) who presented to Edmonton hospitals (Grey Nuns Hospital, Misericordia Hospital, University of Alberta Hospital and Royal Alexandra Hospital, Alberta, Canada) between January 2008 and December 2012 were included. Patients who did not reside in Alberta were excluded, as their charts were not available, as well as those who were deemed palliative at the time of their initial operation. Patients with end stage renal disease had also been excluded from the initial tissue bank (8). All of the patients provided a urine sample following the diagnosis of colorectal cancer, but prior to any operation, chemotherapy or radiation. Ethical approval was obtained from the Health Research Ethics Board at the University of Alberta, and patients provided informed consent.

Charts were reviewed for chemotherapy regimens, delays or discontinuation of chemotherapy, chemotherapy dose reductions, hospitalizations during chemotherapy treatment, cancer progression, cancer recurrence and five-year survival. Patients who had cancer detected while still on chemotherapy were deemed to have progression, whereas patients who had cancer detected on post-treatment surveillance were deemed to have recurrence.

Urine samples were thawed at room temperature. A 1:10 ratio of internal standard (2,2-dimethyl-2-silapentane-5-sulfonate) was added, following which the urine sample was pH adjusted to 6.7–6.8 using 1.0 M HCl or 1.0M sodium hydroxide buffers as required. The urine was pipetted into a Wilmad 528-pp 4-inch nuclear magnetic resonance (NMR) tube (Wilmad, Buena, NJ, USA).

Metabolite spectra were collected using a 600 MHz NMR spectrometer (Oxford Instruments, Abingdon, Oxfordshire, UK) and a VNMRS two-channel console (Varian Inc.; Agilent Technologies Inc., Santa Clara, CA, USA) that was running VNMRJ software version 2.2C (Agilent Technologies, Inc.) on a RHEL 4 (Red Hat) host computer. The spectrometer was equipped with an HX probe and Z-axis gradients. The first increment of a 2D-¹H, ¹H-NOESY pulse sequence acquired the 1H-NMR data and suppressed the solvent signal. Each run used a 100 ms mixing time along with a 990 ms pre-saturation (~80 Hz γB₁). Spectra were collected at 25°C for a total of 32 scans over a period of 3.5 min.

Quantification of 69 validated and previously confirmed metabolites from the spectra was completed using targeted profiling with Chenomx NMR Suite v7.0 software (Chenomx Inc, Edmonton, Alberta, Canada). Quantification was completed by one individual and verified by a second; both individuals were blinded to the provenance of the samples.

The open source software R was used for all analyses (9). Several prediction tasks were examined separately, using 10-fold cross-validation. For each of these analyses, 12 metabolites were removed from the entire dataset as these were previously found to be inconsistently quantified (10). Metabolites with nil values were replaced with half the value of the lowest concentration for that metabolite. The prediction was done using four different machine learning algorithms: least absolute shrinkage and selection operator (LASSO), Support Vector Machine (SVM), Decision Trees and Random Forest. The metabolite concentrations were also log transformed and urea-normalized. Predictive performance results for each prediction task were compared using area under the receiver operating characteristic (ROC) curve (AUC).

Creation of a predictor for each task was performed as described previously (11). For each prediction task (e.g., treatment delay, hospitalization, cancer recurrence), the performance of each combination of machine learning algorithms and data processing was examined (log-transformation/urea-normalization). In 10 fold cross-validation, nine-tenths of the urine samples from each group (with and without the clinical indication being predicted) were randomly assigned to the un-blinded training set. This un-blinded training set was used to generate the metabolite profile that was diagnostic for this prediction task. The remaining one-tenth of the urine samples from each group (with and without the clinical indication) formed the blinded testing set. This blinded testing set was used to validate the metabolite profile diagnostic for the feature of interest. This was repeated 10 times, so that each one-tenth split of the data set acts as the testing set once.

After applying the exclusion criteria, 62 patients remained. As presented in Table I, 34 were male (54.8%) and the age range was 44–87 years with a median age of 66. The primary site of cancer was the colon in 41 patients, and rectum in 22 patients (1 patient had synchronous colonic and rectal lesions).

Of the 62 patients whose charts were reviewed, 45 received chemotherapy. Of the 17 who did not receive chemotherapy, 5 declined for personal reasons, 5 were not offered chemotherapy due to medical co-morbidities, 3 were not offered chemotherapy due to post-operative complications and 4 had no reason given for not receiving chemotherapy.

Of the 45 patients who received chemotherapy, 13 completed their course with no delays, 30 had a delay in treatment and 2 patients had no data. Reasons for treatment delay included reactions to chemotherapy, other medical conditions and patient choice. Reactions to chemotherapy included neutropenia, diarrhea, hand and foot syndrome, mucositis and peripheral neuropathy. Other medical conditions included cellulitis, hip fracture, rib fracture, central line infection and eye surgery. One patient chose to delay treatment to go on vacation.

Of the 45 patients who received chemotherapy, in our predictor for hospitalization, 7 patients required hospitalization, 37 did not and 1 did not have data recorded. Reasons for hospitalization included cellulitis, dehydration, diarrhea, mucositis, SVC obstruction and a hip fracture.

Twenty-three of the 45 patients receiving chemotherapy had dose reductions, 19 did not and 3 had no data recorded. Reasons for dose reduction included hand and foot syndrome, diarrhea, nausea and vomiting, neuropathy, facial rash and reduced creatinine clearance.

Three patients changed chemotherapy type due to drug intolerances, and 5 changed chemotherapy type due to disease progression. No regime change was required in 35 patients, and for 2 patients there was no data recorded.

Five patients had cancer progression while on chemotherapy, and 14 had a recurrence of cancer identified on post-treatment surveillance. Eight patients had no data collected. Recurrences or progressions were detected with colonoscopy, sigmoidoscopy, computed tomography scan, magnetic resonance imaging or physical exam.

At the time of data collection, 5 patients had not reached 5 years since the date of their surgery. Of the 57 patients for which 5 years had elapsed since surgery, 36 were still alive and there were 21 mortalities.

Individual predictors were created for each variable as outlined in the methods section. Predictors were created for 5-year survival, cancer recurrence, chemotherapy dose reduction and treatment delay. The sample size was too small for predictors to be created for hospitalization, regime change and cancer progression.

Each predictor used 4 different machine-learning algorithms (LASSO, SVM, decision trees and random forest) for log-transformed and urea-normalized concentrations of the measured metabolites, for a total of 8 AUC values for each predictor. The AUC values are presented in Tables II and III.

For 5-year survival, the most effective predictor had an AUC of 0.612. The AUC values for other variables included: Cancer recurrence, 0.650; chemotherapy dose reduction, 0.542; and treatment delay, 0.750.