Machine Learning Based Prediction of Distant Metastasis in Rectal Cancer Patients with Complete Response following Neoadjuvant Chemoradiotherapy

Abstract

Distant metastasis (DM) is the leading cause of disease recurrence in rectal cancer patients with clinical or pathological complete response (CR) following neoadjuvant chemoradiotherapy (nCRT). Identifying these patients may help to tailor the treatment approach. The aim of this study was to predict the rectal cancer patients with high risk of DM following CR with nCRT, using machine learning (ML), based on patient, disease and treatment related clinical factors. Data of the rectal cancer patients treated with nCRT at three academic centers between 2010-2022 were collected. Patients with secondary malignancies were excluded. 157 patients with CR were included in the analysis. Median age was 63 (31-85). Female/male ratio was 59/98. Tumor location was proximal-, mid- and distal-rectum in 9, 65 and 83 patients, respectively. cT stage was T2, T3 and T4 in 14, 120 and 23 patients, respectively. 122 patients had cN (+). Median tumor and regional lymph nodes doses were 50.4 (45-56) Gy and 45 (45-50.4) Gy, respectively, in 25-28 fractions, concurrently with capecitabine. 63 patients had clinical CR and were followed up without surgery, and 94 had pathological CR. Median RT-surgery interval was 11 (3.4-59) weeks. 41 patients were managed with total neoadjuvant treatment and 52 received adjuvant systemic treatment. Median follow up time was 37 (3-143) months. DM was detected in 18 patients at median 14 (2-87) months and local tumor recurrence/progression was seen in 9. A ML model was used to predict DM based on patient, disease and treatment related 37 clinical factors. Statistical analysis was conducted using Python (v.3.7). P<0.05 was accepted as statistically significant. Multiple Linear Regression was used to analyze the relationship between DM and independent continuous variables. Chi-square test was used to analyze the difference between the observed frequency distribution of a categorical variable and the expected frequency distribution. Confusion matrix algorithm was used to evaluate the performance of the model. Logistic regression algorithm (LRA) was selected as the best performing algorithm for the ML model that used 37 variables from 157 patients. The model was trained using 75% of the data and the remaining 25% was used for testing. CEA, ratio of post- to pre-RT tumor SUVmax, tumor dose, RT-surgery interval, pre-RT weight loss, smoking history, cN status, and gender were significant parameters to predict DM, based on the LRA of the ML model. Accuracy score of the model was 92% and AUC of the ROC was 74%. Precision (positive predictive value) and recall (sensitivity) of the model were 67% and 50%, respectively. Machine learning has a high rate of correctly classifying patients and good performance in predicting those with DM. Further studies with larger patient numbers are needed.