To the best of our knowledge, this is the first study to use AI to identify high-risk factors for stage II CC and construct a predictive model for this subgroup of patients. Prediction One is a software tool that allows researchers and clinicians perform predictive analytics without training in machine learning or programming skills and is thus well-suited for busy clinical surgeons. Implementing this methodology in the current study allowed us to obtain a high AUC value using everyday clinicopathological factors and without considering special factors.
The linear models with objective and explanatory variables that are currently commonly used within medical statistics are easy to interpret but do not have high prediction accuracy. In contrast, Prediction One enabled us to create a non-linear model using a gradient-boosting tree and neural network, and this methodology led to highly accurate predictions [9]. Moreover, there are high hopes for applying AI in medicine owing to its strong predictive performance. However, because AI uses deep learning techniques to learn from vast amounts of data and derive answers ‘autonomously’, there are problems/concerns regarding humans not understanding the ‘thought process’ behind the results; this is termed the ‘black box’ problem [11]. When AI is used in medicine, it is necessary to comprehensively delineate the rationale behind AI assessments. Moreover, the concept of ‘explainable AI’ [12] has recently been introduced in medicine. Prediction One is considered a ‘white box’ AI methodology because it can calculate IOV values to identify which factors contribute to prediction probabilities and because this methodology provides highly interpretable results.
According to National Comprehensive Cancer Network guidelines [5], T4 lesions, poor differentiation, penetration/obstruction, and LN under-sampling are high-risk factors for stage II CC. Moreover, the European Society for Medical Oncology guidelines [13] list several additional high-risk factors, including venous invasion and high CEA levels. Hence, guidelines published to date demonstrate the wide variety of high-risk factors and the lack of consensus within the current literature. In addition, most subgroup analyses of stage II CC conducted within large clinical trials in Europe and the United States [14,15] showed no additional benefits of AC for stage II CC. This indicates that the high-risk factors described in the current guidelines lack scope and definitiveness, and comprehensively validated methods for predicting high-risk factors are urgently needed. In the present study, only preoperative CEA levels, venous invasion, and obstruction were selected from among the evaluated IOV values.
Regarding bowel obstruction, Sabbagh et al. [16] examined 504 CC cases and reported that the 5-year OS for stage II CC patients who received AC was 92.1% (95% [confidence interval] CI, 86.9–97.6), whereas that for-those who did not receive AC was 80.1% (95% CI, 72.3–88.8). Regarding elevated preoperative CEA levels and venous invasion, Matsuda et al. [6] found that these factors affected DFS in a multivariate Cox regression analysis conducted within a randomized controlled trial sub-analysis. Moreover, with regard to predicting CC prognoses, no reports have examined the superiority of analyses using normal linear models compared with AI analyses using IOV. However, we note that the prediction model constructed in this study had a high AUC.
In further evaluations, we defined the three factors with high IOV values identified in the preliminary phase of the study as high-risk factors and found that stage II CC patients with ≥ 2 of these factors had significantly worse DFS. This finding is similar to the 58% DFS observed in a surgery alone group in a previous report on stage III CC, for which the usefulness of AC has already been established [17]. This suggests that having ≥ 2 of the aforementioned risk factors (i.e. from among identified factors of preoperative CEA, venous invasion, and obstruction) could be considered to define high-risk stage II CC that may be indicated for postoperative AC, although our findings need to be validated within future larger-scale research efforts.
Recently, minimally invasive surgeries, including laparoscopic surgery and robotic surgery, have become mainstream for treating CC; these surgical techniques provide patients with good postoperative quality of life (QOL). However, Grade 3 or higher adverse events were observed in 13.7% of stage II CC patients who received AC using oral tegafur-uracil [6] and in 10–24% of patients who received oxaliplatin-based AC [18], which indicates that patients’ postoperative QOL can be affected because of these adverse events. These findings also demonstrate the importance of stratifying high-risk stage II CC.
This study had several limitations. First, we conducted this single center study with a retrospective design, and deep learning analysis was performed in a small sample. Second, because we only performed five-fold cross-validation without external validation, the possibility of overfitting cannot be ruled out. Third, we did not examine tumor budding, fibrotic cancer stroma, microsatellite instability, or loss of 18q heterozygosity [19], all of which have recently been reported as promising factors associated with high-risk stage II CC. In the future, we recommend that a more robust predictive model for high-risk stage II CC needs to be established for a multicenter study enrolling a large sample size and with external validation and that this model should include these promising factors.
In conclusion, we constructed a new predictive model for high-risk stage II CC with high validity and high probability using an auto-AI tool, Prediction One. Our findings, which should be confirmed within future more highly powered research efforts, suggest that patients with ≥ 2 of the following factors are high-risk stage II CC patients who are likely to benefit from AC: preoperative CEA levels > 5.0 ng/mL, venous invasion, and obstruction. Our findings guide future research directions and directly inform medical guidelines as well as clinical decision-making.