Development of a prognostic prediction model for tumor recurrence or progression in gastrointestinal stromal tumors: integrating inflammatory blood indices with 5 mm2 versus 50 high-power field mitotic counts
Article information
Abstract
Background
The behavior of gastrointestinal stromal tumors varies according to tumor size, location, mitotic rate, and rupture. Traditional mitotic counting in 50 high-power fields (HPF; 5.3 mm2) may overestimate progression risk, as 5 mm2 corresponds to 20–25 HPF under modern microscopy. However, the 50-HPF method remains widely used. This study evaluated the impact of two mitotic counting methods on risk classification and developed a prognostic model combining inflammatory biomarkers with clinicopathological factors.
Methods
We retrospectively analyzed 150 patients who underwent resection at Songklanagarind Hospital between 2009 and 2021. Patients who had received neoadjuvant therapy or had a second primary cancer were excluded. Mitosis was evaluated using both methods. Clinicopathological and serum biomarker data were collected. The primary outcome was 5-year disease-free or progression-free survival. The model was developed using forward stepwise variable selection based on the time-dependent area under the curve and internally validated by bootstrapping.
Results
The 50-HPF method overestimated the Armed Forces Institute of Pathology classification in 14.2% of cases. A model incorporating tumor size, location, rupture, mitotic count within 5 mm2, and systemic immune-inflammation index achieved the highest area under the curve of 0.939 in both training and validation sets, outperforming previous models. The 5-mm2 method demonstrated superior performance to the 50-HPF method across all models (0.939 vs. 0.919 for the proposed model).
Conclusions
Mitotic counting in 50 HPF overestimated risk classification in certain cases. Incorporating systemic immune-inflammation index and using a 5 mm2 mitotic count enhanced prognostic model performance.
INTRODUCTION
Gastrointestinal stromal tumors (GISTs) are the common mesenchymal tumors of the digestive tract, predominantly originating in the stomach [1]. Their behavior exhibits wide variability in progression rates, depending on tumor size, location, mitotic rate, and rupture status [2-4]. Several risk classification systems have been proposed to predict disease progression and guide adjuvant therapy, including the Armed Forces Institute of Pathology (AFIP) [2] and the modified National Institutes of Health (NIH) criteria [3].
The AFIP criteria, one of the most widely used systems proposed by Miettinen et al., were initially developed based on mitotic counts in 50 high-power fields (HPF), corresponding to an area of 5.3 mm2 [5,6]. This approach was established in an era when 50 HPF approximated 5 mm2. However, because the field area depends on the ocular field number, an area of 5 mm2 in modern microscopes corresponds to only 20–25 HPF. As a result, mitotic counting across 50 HPF, which encompasses a substantially larger area, may yield higher mitotic counts and consequently inflate risk classification [7]. To minimize this variability, recent guidelines have recommended mitotic counting within a standardized 5 mm2 field area instead of 50 HPF [8-12]. In 2023, Campora et al. [13] demonstrated that the 50-HPF method could overestimate progression risk by 10%–41%, depending on the field diameter. Such misclassification may have important clinical implications, particularly in decisions regarding adjuvant therapy. However, their survey revealed that more than half of pathologists still use the traditional 50-HPF method.
Recent studies have highlighted the prognostic value of blood indices reflecting inflammatory and nutritional status, including the systemic immune-inflammation index (SII), platelet-to-lymphocyte ratio (PLR), neutrophil-to-lymphocyte ratio (NLR), prognostic nutritional index (PNI), and hemoglobin, albumin, lymphocyte, and platelet (HALP) score in various cancers [14-19]. Incorporating these indices can enhance the prognostic performance of established risk classification systems [20-23]. However, as mitotic activity is a key component of these systems, variability in mitotic counting methodology may influence their prognostic performance. In these studies, mitotic activity was assessed using the conventional 50-HPF method, which may have affected standardization across different microscopic settings. Therefore, there remains a need to evaluate the prognostic impact of integrating these biomarkers with standardized mitotic assessment within a prognostic model, and to identify the optimal combination of biomarkers and mitotic counting methods for risk prediction.
Accordingly, this study aimed to evaluate the impact of the 5 mm2 and 50-HPF mitotic counting methods on risk classification. We further developed a prognostic model to assess their effects on risk prediction and to enhance model performance by incorporating systemic inflammatory and nutritional biomarkers as complementary variables. The predictive performance of this model was compared with that of established risk classification systems. The model is intended to remind pathologists of the appropriate mitotic counting method and to guide clinicians in surveillance planning for high-risk patients.
MATERIALS AND METHODS
Study design, population, sample size, and data collection
This retrospective cohort study included patients diagnosed with GISTs who underwent surgery between January 2009 and December 2021 at Songklanagarind Hospital, a single tertiary care center in Songkhla, Thailand. Patients were excluded if they had received neoadjuvant therapy or had two primary cancers.
Variables, including age, sex, adjuvant treatment, tumor site, date of surgery, date of first recurrence or metastasis, last follow-up, and intraoperative parameters, such as tumor rupture status, were retrieved from the Hospital Information System of Songklanagarind Hospital and the Digital Innovation and Data Analytics database, Faculty of Medicine, Prince of Songkla University.
Pathological assessment
Archived formalin-fixed paraffin-embedded blocks were retrieved from the Department of Pathology, Prince of Songkla University. Each tissue block was sectioned at a thickness of 3 µm for hematoxylin and eosin staining. All samples were reexamined by a pathologist who was blinded to clinical history and outcomes. Additionally, 10% of cases were randomly selected for mitotic count assessment by a second pathologist. In cases of mitotic grade discordance, both pathologists jointly reviewed the slides and reached a consensus.
All pathological features were evaluated using a light microscope (field number 22, Nikon ECLIPSE Ci, Tokyo, Japan). Mitotic activity was assessed in two ways: within a hotspot area of 5 mm2 (equivalent to 21 HPF using a 40× objective lens) and across 50 HPF. A mitotic cut-off of 5 was used to classify tumors into low (≤5) and high (>5) mitotic grade categories [24]. Tumor size was determined based on the greatest diameter described in the gross section of the pathology report. Histologic subtypes were categorized as spindle cell, epithelioid, or mixed. Additional features, including tumor necrosis, ulceration, mucosal invasion, muscle invasion, and serosal involvement, were recorded as present or absent.
Laboratory assessment
Preoperative serum laboratory data, including albumin levels and complete blood counts, were collected within 2 months prior to surgery and used to derive several inflammation- and nutrition-based indices. The PNI was calculated as follows: 10 × serum albumin (g/dL) + 0.005 × total lymphocyte count (/mm3). The HALP score was computed as hemoglobin (g/L) × albumin (g/L) × lymphocyte count (/L)/platelet count (/L). The SII was calculated as follows: platelet count (/L) × neutrophil count (/L)/lymphocyte count (/L). Other derived markers were calculated based on direct ratios, including the PLR, NLR, and monocyte-to-lymphocyte ratio (MLR).
Statistical analysis
Clinicopathological characteristics, intraoperative findings, and laboratory parameters were summarized using means with standard deviations or medians with interquartile range (IQR) for continuous variables and percentages for categorical variables. As no established cut-off values exist for blood indices, optimal thresholds were determined using the Euclidean distance method, selecting the receiver operating characteristic (ROC) point closest to the ideal coordinate [0,1] [25,26].
Disease-free survival (DFS) was defined as the interval from surgery to the first recurrence or metastasis in patients with stage I–III disease. For patients with stage IV disease, progression-free survival (PFS) was defined as the interval from surgery to disease progression, indicated by locoregional recurrence or new metastatic lesions. For patients without recurrence or progression at the time of analysis, survival data were censored at the last follow-up (March 31, 2025), which marked the end of the study period. To address missing data, multiple imputation by chained equations [27] was performed before area under the curve (AUC) estimation.
The associations between variables and prognosis were evaluated using Kaplan-Meier survival analysis and Cox proportional hazards regression. To reduce heterogeneity, survival outcomes were analyzed separately, with DFS for stage I–III patients and PFS for stage IV patients. Variables that were significant in univariate Cox regression analysis were included in the multivariable model.
We developed a model to predict DFS/PFS. All available cases, including stages I–IV, were used for model development. Forward stepwise variable selection based on the time-dependent AUC was used for model construction. The initial model included four established prognostic factors identified in the literature [3,24]: tumor site, tumor size, mitotic count, and tumor rupture. Additional variables were subsequently incorporated to enhance the 5-year time-dependent AUC. The final prognostic models were compared with previously established scoring systems, including the AFIP criteria, modified NIH criteria, and the 8th edition of the American Joint Committee on Cancer (AJCC) staging system, based on AUC performance. AUCs were computed separately for models utilizing mitotic counts measured per 5 mm2 and per 50 HPF to compare the prognostic performance of the two counting methods.
Model calibration was evaluated using three complementary approaches: calibration curves, the Hosmer-Lemeshow goodness-of-fit test [28], and the Brier score [29]. Internal validation was performed using bootstrapping. Statistical analyses were performed using R ver. 4.4.2 (R Foundation for Statistical Computing, Vienna, Austria), and statistical significance was set at p < .05.
RESULTS
Patients’ characteristics
A total of 214 patients were initially eligible. After excluding 40 patients with two primary cancers and 24 patients who received neoadjuvant therapy, 150 patients were included in this study. The clinicopathological characteristics of the patients are summarized in Table 1. Fifty-seven patients (38.0%) were male and 93 (62.0%) were female, with a mean age of 61.5 ± 13.4 years. Most tumors were unifocal, comprising 147 cases (98.0%). The stomach was the most common primary tumor site, accounting for 82 cases (54.7%). Tumor rupture occurred in 23 patients (15.3%). The median tumor size was 7.0 cm (IQR, 4.0 to 11.9 cm). According to the 8th edition of the AJCC staging system, 67 cases (44.7%) were stage I, 27 (18.0%) were stage II, 43 (28.7%) were stage III, and 13 (8.7%) were stage IV. Eighteen patients (12.0%) received adjuvant imatinib. The median mitotic counts were 3 (IQR, 1 to 12) and 6 (IQR, 2 to 21) when assessed in 5 mm2 and 50 HPF, respectively.
Clinicopathological characteristics of patients with resected gastrointestinal stromal tumors (n = 150)
Impact of mitotic counting method on AFIP risk classification
Using a mitotic cut-off of 5, mitotic counting in 50 HPF resulted in a higher grade in 22 cases (14.6%) compared with assessment in 5 mm2, whereas the remaining 128 cases retained the same mitotic grade, with 72 classified as low grade and 56 as high grade.
According to the AFIP criteria, risk category upgrading occurred in 20 cases (14.2%) when mitoses were counted using the 50-HPF method compared with the 5 mm2 assessment (Table 2). Specifically, six cases (4.3%) were reclassified from very low to moderate risk, 11 cases (7.8%) from low to high risk, and three cases (2.1%) from moderate to high risk. Nine cases were unclassifiable because of insufficient data based on the AFIP criteria.
AFIP risk classification based on two mitotic counting methods
Patient survival according to the mitotic counting method
The median follow-up duration was 61.3 months (IQR, 32.9 to 67.2 months). Among the 39 patients (26.0%) who experienced recurrence, 10 (6.7%) had local recurrence, 35 (23.3%) had metastasis, and six (4%) had both. Approximately 25% of patients experienced recurrence or progressive disease by 54.8 months (95% confidence interval [CI], 41.3 to 71.8 months). The median time to recurrence could not be estimated, as the recurrence rate did not reach 50% during the follow-up period.
In patients with stage I–III disease, those with a high mitotic grade demonstrated significantly worse DFS than those with a low mitotic grade, regardless of the counting method used (p < .001) (Fig. 1A, B). Survival analysis for PFS in stage IV patients was not feasible due to the limited sample size (n = 13).
Kaplan-Meier curves of disease-free survival in stage I–III patients. (A) Mitotic count per 5 mm2. (B) Mitotic count per 50 high-power fields.
Kaplan-Meier survival analysis stratified by blood indices
Patients with high SII values had significantly worse DFS than those with low SII values (p = .006) (Fig. 2A). At 5 years, the estimated probability of recurrence was 49.2% (95% CI, 22.4% to 66.7%) in the high SII group versus 17.3% (95% CI, 8.8% to 25.1%) in the low SII group.
Kaplan-Meier curves of disease-free survival in stage I–III patients. (A) Systemic immune-inflammation index (SII). (B) Platelet-to-lymphocyte ratio (PLR). (C) Monocyte-to-lymphocyte ratio (MLR). (D) Neutrophil-to-lymphocyte ratio (NLR).
In addition to the SII, high PLR and MLR were also significantly associated with worse DFS, as demonstrated in Fig. 2B and C, respectively. In contrast, NLR, PNI, and HALP were not significantly associated with DFS (Fig. 2D for NLR; data for other indices not shown).
Univariate and multivariable Cox regression analyses
Univariate and multivariable Cox regression analyses for DFS in stage I–III patients are presented in Table 3. The presence of tumor rupture, tumor size >5 cm, mitotic count >5 per 5 mm2 and per 50 HPF, as well as high SII, high PLR, and high MLR, were significantly associated with worse DFS in univariate analysis. In multivariable analysis, only tumor size, mitotic count, and high PLR remained independent prognostic factors for DFS.
Univariate and multivariable Cox regression analyses of DFS in stage I–III patients
Model performance of prediction models
The time-dependent AUC of the model comprising mitotic count (≤5 vs. >5 per 5 mm2 or per 50 HPF), tumor size (≤5 vs. >5 cm), tumor site (gastric vs. non-gastric), and tumor rupture (presence vs. absence) was 0.920 (95% CI, 0.867 to 0.974) and 0.910 (95% CI, 0.855 to 0.966) when mitotic count was assessed per 5 mm2 and per 50 HPF, respectively.
The addition of SII (≤1,358.6 vs. >1,358.6) provided the greatest improvement in predictive performance, increasing the AUC to 0.939 (95% CI, 0.894 to 0.983). The incorporation of PLR (≤16.5 vs. >16.5) also improved model performance, yielding an AUC of 0.937. The AUC values for models incorporating other variables are presented in Supplementary Table S1. Model performance, evaluated separately using mitotic counts per 5 mm2 and per 50 HPF, is summarized in Table 4, along with corresponding time-dependent AUCs and 95% CIs.
Discriminative performance of the proposed and previous models for 5-year DFS/PFS prediction
The final prognostic model comprised mitotic count, tumor size, tumor site, tumor rupture, and SII. To benchmark prognostic discrimination, it was compared with previously established models, including the AFIP criteria, modified NIH criteria, and the 8th edition AJCC staging system, both with and without SII inclusion (Table 4). The proposed model demonstrated superior discriminative ability, achieving the highest AUC of 0.939 (95% CI, 0.894 to 0.983) when using mitotic count per 5 mm2 and 0.919 (95% CI, 0.867 to 0.971) when using mitotic count per 50 HPF. The ROC curves comparing the proposed model with previous models based on mitotic counts per 5 mm² are shown in Fig. 3. The AFIP risk grouping applied to 141 patients, as nine cases could not be classified because of undefined categories within the criteria. In a subgroup analysis of stage I–III patients, the proposed model demonstrated similarly high discriminative performance, with AUCs of 0.946 (95% CI, 0.902 to 0.991) and 0.923 (95% CI, 0.870 to 0.977) for mitotic counts assessed per 5 mm2 and 50 HPF, respectively.
Receiver operating characteristic curves comparing the proposed model with previous models based on mitotic counts per 5 mm2. AUC, area under the curve; AFIP, Armed Forces Institute of Pathology; NIH, National Institutes of Health; AJCC, American Joint Committee on Cancer.
Fig. 4 illustrates the calibration curve for 5-year DFS/PFS. The model tended to overestimate risk within the predicted probability range of 0.2–0.6. The Hosmer-Lemeshow test yielded a p-value of .913, indicating good model calibration, with a Brier score of 0.102.
Calibration curve comparing predicted probabilities with observed outcomes for 5-year disease-free survival/progression-free survival.
Table 4 presents the results of internal validation. The proposed model achieved its highest discriminatory performance, with an AUC of 0.939 (95% CI, 0.889 to 0.979), when mitotic count was evaluated per 5 mm2. Across both the original model and the internal validation set, mitotic count measured per 5 mm2 consistently yielded a higher time-dependent AUC than the 50-HPF method.
DISCUSSION
Mitotic counting using the conventional 50-HPF method resulted in overestimation of the AFIP risk classification in certain cases. Our proposed model, which incorporated tumor size, location, rupture status, mitotic count within a 5-mm2 area, and SII, achieved the highest time-dependent AUC for DFS/PFS compared with previously established models. Furthermore, mitotic counting within a 5-mm2 area consistently yielded better prognostic performance than the 50-HPF method across all models.
Using the 50-HPF method for mitotic counting resulted in overestimation of AFIP risk classification in certain cases, consistent with Campora et al. [13], who reported that mitotic assessment over 50 HPF could overestimate risk in 10%–41% of cases, depending on the microscope field diameter. The AFIP criteria were originally established using a mitotic count threshold of five per 50 HPF, corresponding to an area of approximately 5.3 mm2 [5,6]. With modern microscopes, the area covered by 50 HPF has nearly doubled, as 5 mm2 now corresponds to only 20–25 HPF. Consequently, counting mitoses across 50 HPF may overestimate mitotic activity, resulting in an upward shift in mitotic grade and AFIP risk classification. In our study, we observed a similar upward shift when using the 50-HPF method, consistent with the AFIP criteria. Nevertheless, Campora et al. [13] found that 64.5% of 110 pathologists still relied on the 50-HPF method for mitotic counting. Despite variation in counting methods, our Kaplan-Meier analysis demonstrated that a mitotic rate of ≥5 was significantly associated with a poorer prognosis than a mitotic rate <5, consistent with previous studies [5,30,31]. Our findings further indicate that the method of mitotic assessment is a key determinant of prognostic performance. Consistently, models based on mitotic counts assessed per 5 mm2 demonstrated superior performance compared with those based on counts per 50 HPF across all models. This observation aligns with current clinical guidelines recommending a standardized field area to minimize variability in field diameter [8–12].
Recent studies have demonstrated the prognostic value of nutritional and inflammatory blood indices, including PNI, HALP, SII, PLR, and NLR, across various cancers [14-17,32,33]. Similar findings have been reported in GISTs [21,23,34-38]. However, these indices have not yet been incorporated into routine clinical practice. In our study, univariate analysis showed that high SII, PLR, and MLR were significantly associated with poorer DFS, consistent with previous findings [35-37,39-42]. The underlying mechanisms linking these indices to prognosis remain unclear. A plausible explanation involves the multifaceted roles of inflammatory cells in tumor progression. Platelets, activated by tumor-derived factors, release cytokines, including transforming growth factor-β and vascular endothelial growth factor, promoting epithelial–mesenchymal transition, angiogenesis, and metastasis [43-45]. Neutrophils facilitate tumor spread via extracellular matrix degradation and growth factor release [46,47]. In contrast, lymphocytes, particularly cytotoxic T cells, exert antitumor effects, and higher tumor-infiltrating lymphocyte levels correlate with better outcomes [48]. Monocytes promote tumor growth through angiogenesis, matrix remodeling, and immune suppression [49]. Accordingly, elevated SII values driven by increased platelets and neutrophils or decreased lymphocytes indicate a higher risk of disease progression. Similar associations were observed for PLR and MLR. In addition, SII has been reported to be significantly correlated with circulating tumor cells, which are associated with recurrence and distant metastasis [18]. Incorporating SII into the prognostic model may therefore provide additional information for risk stratification and potentially aid in decisions regarding surveillance intensity. Patients with elevated SII may warrant closer follow-up compared with those with lower SII.
The proposed prognostic model incorporated mitotic count per 5 mm2, tumor size, site, rupture, and SII. It achieved the highest AUC compared with previously established models [2,3,50], both with and without SII. This finding was consistent for both the original and validation sets. In 2015, Goh et al. [20] also reported that integrating PLR or NLR with existing models using either the AFIP or NIH criteria enhanced prognostic discrimination, achieving the highest AUC of 0.866. In a similar manner, the inclusion of SII in our model further enhanced predictive performance. Notably, SII was incorporated as a complementary biomarker to improve model discrimination rather than as a primary prognostic determinant. SII, which is derived from three components—platelets, neutrophils, and lymphocytes—may provide a more comprehensive reflection of host inflammation and immune response to the tumor than indices based on two variables, such as PLR, NLR, and MLR. Accordingly, it has been shown to have superior prognostic performance in several malignancies [51,52]. However, in our cohort, its advantage over PLR was modest. This may be explained by the shared components of these indices, suggesting that the additional contribution of neutrophils may be limited in certain clinical settings, resulting in comparable predictive performance between SII and PLR.
This study emphasizes the importance of evaluating mitotic counts within a standardized 5 mm2 area rather than across 50 HPF to avoid risk overestimation, which may lead to overtreatment. Adopting our model, which incorporates mitotic count per 5 mm2 and SII, could enable more accurate patient stratification. This approach may help reduce unnecessary adjuvant therapy in patients reclassified into lower-risk categories by avoiding the overestimation inherent in the 50-HPF method, while ensuring that truly high-risk patients are correctly identified and treated.
This study had some limitations. First, it was a retrospective, single-center study, and the proposed model has not yet been externally validated. Moreover, the data-driven dichotomization of continuous variables may increase the risk of overfitting. Future studies are warranted to externally validate this model. Additionally, research on dynamic changes in SII and other inflammatory markers during follow-up may provide valuable insights for treatment monitoring and early recurrence detection.
In conclusion, mitotic counting using the conventional 50-HPF method may overestimate AFIP risk classification in certain cases. In addition, mitotic count assessment within a standardized 5 mm2 area provides superior prognostic performance, underscoring its importance in risk stratification. A prognostic model incorporating mitotic count within a 5 mm2 area, along with tumor size, tumor location, rupture status, and SII, demonstrated improved predictive performance for 5-year DFS/PFS compared with existing models.
Supplementary Information
The Data Supplement is available with this article at https://doi.org/10.4132/jptm.2026.06.15.
Notes
Ethics Statement
This study was approved by the Human Research Ethics Unit of the Faculty of Medicine, Prince of Songkla University (approval number: REC.67-560-5-1) in accordance with the Declaration of Helsinki and its later amendments. Formal written informed consent was not required, with a waiver granted by the appropriate Institutional Review Board.
Availability of Data and Material
The datasets generated or analyzed during the study are available from the corresponding author on reasonable request.
Code Availability
Not applicable.
Author Contributions
Conceptualization: WT. Data curation: WT, PT, TI. Formal analysis: WT. Funding acquisition: WT. Investigation: WT. Methodology: PT, TI. Project administration: WT. Resources: WT. Supervision: PT, TI. Validation: PT. Visualization: WT. Writing—original draft: WT. Writing—review & editing: PT, TI. Approval of final manuscript: all authors.
Conflicts of Interest
The authors declare that they have no potential conflicts of interest.
Funding Statement
This study was supported by the Faculty of Medicine, Prince of Songkla University (grant number 67-090-1).
Acknowledgments
This study was supported by the Department of Pathology, Faculty of Medicine, Prince of Songkla University.