In this study, we analyzed a cohort of 210 patients with IP at our center to calculate 1-year, 3-year, and 5-year recurrence free survival rates. We performed Cox regression to identify demographic and clinical features associated with relapse post-IP surgery. Subsequently, we constructed prediction models using various machine learning algorithms and visualized feature importance using SHAP. The bee colony summary map of the EST model revealed that bone defects was the most influential factor for postoperative recurrence. Other significant predictors included orbital involvement, smoking, no processing of tumor attachment sites, drinking, mild or moderate dysplasia, skull base involvement, and Krouse T4. To our knowledge, this study is the first to develop a clinical feature-based interpretable machine learning model for predicting recurrence free survival following IP surgery.

It is well-recognized that IP exhibits a high recurrence rate postoperatively. Previous research indicates that the likelihood of no recurrence follows a bimodal distribution over the duration of follow-up, with specific peaks in recurrence times. In this study, 20.00% of cases experienced a recurrence within an average follow-up period of 44.26 months. The median time to recurrence was 35.50 months. The probabilities of no recurrence at 1, 3, and 5 years were 97.00%, 83.61%, and 75.02%, respectively. These findings are largely in alignment with existing literature [20, 21].

Lifelong follow-up is clinically recommended [10], yet patient compliance is low due to high time and economic costs. Consequently, constructing a predictive model to assess the likelihood of postoperative recurrence is crucial for guiding follow-up strategies. While numerous studies have explored the risk factors for recurrence after IP surgery, few have developed predictive models. Machine learning algorithms, which can identify nonlinear correlations, demonstrate superior performance compared to traditional statistical methods [22]. In addressing survival analysis, many studies classify outcomes at different time points as binary events, which can misrepresent the actual survival outcomes over time [23]. To circumvent these limitations, we utilized the scikit-survive package to construct and validate a prognostic model. Our three models performed consistently well with the training set characteristics and maintained accuracy during the first 55 months of the test set before experiencing a significant drop in predictive effectiveness. Conversely, the other models showed consistent predictive performance for the first 30 months. This divergence likely stems from differing risk factors influencing early versus late recurrence after IP surgery. We hypothesize that early recurrence is predominantly determined by the thoroughness of the surgery, whereas late recurrence may be influenced by factors such as low-risk HPV infection [24, 25]. However, the small number of late recurrence cases and potential increases in machine learning algorithm errors could also reduce the AUC, a hypothesis that warrants verification through larger-scale studies.

To elucidate the importance of variables in machine learning models, we introduced SHAP values, which are instrumental in explaining the interpretability of various models. Although age was not identified as a risk factor in the EST model, SHAP values indicated that younger patients (represented by bluer values) were more prone to relapse in both the GBSA and RSF models. This increased risk in younger individuals may be attributed to more invasive tumors or longer follow-up periods. A German cohort study corroborates this perspective [26]. Regarding lifestyle factors, while different studies present varying conclusions on whether smoking promotes recurrence [21, 27], both smoking and drinking were significant promoters of IP recurrence in our EST model. However, hypertension and diabetes were not identified as potential risk factors. In our study, 50 patients with primary surgery in other hospitals had a slightly higher recurrence rate (26.00%) compared to 18.13% among the other 160 patients with that in our hospital, though the difference was not statistically significant. In addition, neither a history of IP nor prior nasal surgery significantly affected postoperative recurrence. In contrast, a study by Oisín Bugter suggested that a history of nasal surgery could predispose patients to recurrence [28]. Furthermore, research from the University of Pennsylvania highlighted significantly higher inflammation levels in the contralateral nasal cavity in IP, suggesting a link to chronic inflammation [29], although our findings did not support a correlation between chronic sinusitis and postoperative IP recurrence. Lisan et al. found that Krouse T3 paranasal sinus IP posed a higher recurrence risk than T2 [30]. In our data, Krouse T4 was more likely to recur than stage 1, likely due to tumors breaching bone barriers and invading adjacent tissues, complicating surgical removal. The most common sites of origin in our study were the maxillary sinus (40.48%) and the nasal cavity (28.57%), with high recurrence rates observed in tumors originating from multiple sites (70.00%) and the frontal sinus (50.00%). Despite the efficacy of DrafIIb and DrafIII surgeries [31], maintaining the integrity of the frontal sinus proved challenging. A study in South Korea found that the site with the highest recurrence rate after maxillary sinus surgery was the lateral wall [32], which was not thoroughly analyzed in our study. Meanwhile, tumor base, piecemeal resection, and bone defects were not risk factors for postoperative recurrence in IP surgery. Prior research has shown that surgery guided by attachment location significantly reduced recurrence risks [33, 34]. Complete resection during initial surgery is critical for preventing recurrence, as suggested by Healy et al., who found that total removal of bone under the tumor attachment site reduces recurrence rates compared to simpler dissection [35]. In our cohort, not performing special treatment on tumor attachment sites significantly increases the risk of recurrence compared to removing the bone at the attachment site. Patients with bone defects, orbital involvement, or skull base involvement were more susceptible to recurrence. The difficulty in achieving precise resection in the complex anatomy of areas like the pterygopalatine fossa, infratemporal fossa, and surrounding orbital and intracranial tissues often leads to residual tumor tissue, increasing recurrence risks. This conclusion aligns with findings by Wang et al. and supports our observations [36].

The study has several limitations. First, it is a single-center retrospective study, and the narrow range of sample selection may limit its generalizability. Second, to counter the low incidence of IP and the imbalance between the relapsed and non-relapsed groups, SMOTE resampling and fivefold cross-validation were employed, which might have influenced the model's performance. Future studies should incorporate a large, multicenter data set to validate our conclusions. Third, although the data set was divided into 80% for training and 20% for testing, our model has not undergone external validation. This aspect is crucial and will be addressed in future research.

In this study, the recurrence rate of IP patients was 20.00%, with a median recurrence time of 35.50 months. Multivariate Cox regression analysis identified mild or moderate dysplasia in pathological biopsies as an independent risk factor for IP recurrence. In addition, we employed machine learning algorithms, specifically the EST model within our ML prognosis framework, demonstrating their suitability for clinical research in rare diseases with small sample sizes. The EST model highlighted five key predictors of IP recurrence, including bone defects, orbital involvement, smoking, no processing of tumor attachment sites, drinking.