Purpose/Objective(s)
Following
neoadjuvant chemo-radiotherapy (nCRT), some rectal cancer patients may
have pathological lymph node (LN) involvement despite complete response
in the primary tumor and should not be candidates for non-operative
management. 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 pLN involvement following nCRT, using machine
learning (ML), based on pre-surgical patient-, disease- and
treatment-related clinical factors.
Materials/Methods
Data
of the rectal cancer patients treated with nCRT between 2013-2023 were
collected. Patients with secondary pelvic malignancies, no LN
dissection, RT-surgery interval > 21 weeks and cLN diameter < 5 mm
were excluded. 143 patients were included in the analysis. Median age
was 58 (26-83). Female/male ratio was 55/88. Tumor (T) location was
proximal-, mid- and distal-rectum in 8, 58 and 77 patients,
respectively. cT stage was T2, T3 and T4 in 9, 122 and 12 patients,
respectively. 89 patients had cLN diameter ≥ 10 (10-34) mm. Except 16
(mucinous: 11, signed ring cell: 5) patients, all had adeno carcinoma
NOS histology. Median pre-RT T and LN SUVmax were 15 (4-46) and 3 (1-30), respectively. Median GTV was 73 (23-370) cm3.
Median T length and thickness were 6 (2-14) cm and 16 (7-48) mm,
respectively. Median T and regional LN doses were 56 (45-56) Gy and 50.4
(45-50.4) Gy, respectively, in 25-28 fractions, concurrently with
capecitabine. Median RT-surgery interval was 11 (4-21) weeks. 48
patients were pN (+). A ML model was used to predict pLN involvement
based on 32 parameters (including age, gender, BMI, smoking, Hb, Hct, T
localization, histology, GTV, and cLN diameter). Logistic regression
model, configured with hyperparameters, exhibited the best performance
with high accuracy and F1 scores after testing various algorithms. The
performance of the model was evaluated using precision (positive
predictive value), recall (sensitivity), F1 score (performance of
classification), accuracy, and ROC AUC value. The model was trained
using 75% of the data and the remaining 25% was used for testing.
Results
Accuracy,
F1 scores and ROC AUC value of the model were 85.71%, 70.59% and 0.78,
respectively. Precision, recall and F1 score of the model are seen in
the Table. Analysis of the confusion matrix revealed
that the model successfully predicts 24 true negative and 6 true
positive cases, with only 1 false positive and 4 false negatives.
Similar accuracy rates on the training (84.76%) and test (85.71%)
datasets were observed. Histology, BMI, Hct, Hb, and weight were the
most significant parameters in the model.
