Definitive Hypofractionated Radiotherapy in Elderly Patients With Non-melanoma Skin Cancer: Predicting Complete Response Using a Machine Learning-based Approach

Abstract

Background/Aim

This study aimed to develop a classification tree model to predict clinical complete response (cCR) after definitive hypofractionated RT in elderly patients with non-melanoma (basal and squamous cell carcinoma) skin cancer based on pre-treatment clinical parameters.

Patients and Methods

Data from elderly patients (aged ≥ 70 years) with histologically proven non-melanoma skin cancer lesion(s) were reviewed. All lesions were treated with definitive hypofractionated RT. A decision tree methodology was applied to define significant pre-treatment parameters influencing the cCR rate. The data set included seven predictor variables (location, histology, T stage, diameter, bleeding, pain, previous treatment) and a dependent variable (cCR). The analysis was conducted using open-source R software. Important explanatory variables were selected using the random forest approach. A classification tree that optimally partitioned patients according to cCR was then built.

Results

A total of 121 patients with 145 non-melanoma skin cancer lesions were identified between January 2015 and December 2023. The cCR rate after treatment was 78.6%. The three most important variables identified were T stage, location and pain at diagnosis. Based on these explanatory variables, patients were partitioned in five groups. Of these groups, three experienced cCR rates ranging from 69% to 88% and two experienced cCR rates lower than 15%.

Conclusion

The proposed classification tree could serve as a valuable tool to guide future research in the field of non-melanoma skin cancer. The decision tree method may be an effective approach to partition elderly patients with non-melanoma skin cancer based on key clinical parameters.

Introduction

Non-melanoma (basal and squamous cell carcinoma) skin cancer is a type of radiosensitive malignancy. Local treatment consists of surgical resection and/or radiotherapy (RT) (1, 2). RT typically involves conventional 1.8-2 Gy fractions administered daily over 5-7 weeks with a cumulative dose ranging from 45 to 70 Gy, based on the treatment purpose (adjuvant or definitive) (1, 2). Definitive RT is commonly used in elderly patients, as various factors (such as comorbidities, use of anticoagulants, cognitive impairment, functional losses, tumor location and patient preference) may preclude surgical management. Hypofractionated RT approaches are often pursued to reduce overall treatment time and enhance patient and family convenience (3).

Based on the encouraging results of the neoadjuvant therapy with cemiplimab, an anti-programmed cell death 1 (PD-1) monoclonal antibody, in the setting of resectable squamous cell carcinoma – showing a pathological complete response of 51%, an objective response rate of 68% and severe or higher adverse event frequency of 18% – immunotherapy is increasingly being investigated as a potential curative approach (4, 5). A shift in treatment strategy is expected in the very near future, identifying factors associated with clinical complete response (cCR) after RT could help guide the most effective treatment strategy.

In this context, we analyzed data from elderly patients with non-melanoma skin cancer who underwent definitive hypofractionated RT using classification trees and random forests. Our aim was to contribute to the existing literature on clinical outcome predictors using machine learning. We hope that this work represents an exploratory step towards personalized treatment selection for elderly patients with non-melanoma skin cancer utilizing pre-treatment clinical parameters.

Patients and Methods

Patient population. Patient data were retrieved from three Italian Radiotherapy Departments participating in this retrospective analysis approved by the Institutional Review Board of the promoting Institution (Rif. 6452 Prot. 0712/2021). Each patient provided informed consent prior to the use of the clinical data for research purposes.

We collected data from elderly patients with histologically proven non-melanoma skin cancer lesion(s) who received RT with curative intent from January 2015 to December 2023. Elderly patients were defined as those aged ≥70 years. The treatment of all lesions involved definitive hypofractionated RT, with the treating radiation oncologist determining the specific protocol to be employed, in accordance with current clinical guidelines.

Data collection and endpoint definition. The collected data included baseline patient and tumor characteristics, including comorbid conditions, Eastern Cooperative Oncology Group (ECOG) performance status (PS), and tumor outcomes. Tumor pain and bleeding grades were assessed at the peak grade before RT, at the end of RT and three months after treatment. The retrospective nature of the study presented inherent limitations, which potentially could have led to missing data. Consequently, we did not analyze acute and late toxicities.

Complete response (cCR) was defined as the complete clinical disappearance of the irradiated at three months after RT, with no residual tumor detectable clinically and/or by imaging. Decision tree methods were used to select important variables for cCR prediction.

Survival times were defined as the time in months from the end of RT to the first event of interest, including death from any cause for overall survival (OS) and/or disease recurrence for progression-free survival (PFS). Patients were censored at their last follow-up date. Locoregional control was also analyzed and was defined on a “per lesion” basis, as relapse in the irradiated treatment volume.

Statistical analysis. Statistical analysis was performed using the open-source RStudio version 0.98.1091 . Standard descriptive statistics were used to evaluate the distribution of each variable. Continuous variables were reported as mean±standard deviation (SD) and categorical variables as frequencies or percentages. OS and PFS were estimated using the Kaplan-Meier method.

In addition to these standard statistical methods, a machine-learning-based methodology was applied to define significant pre-treatment variables that could predict cCR.

The data set contained seven predictor variables related to each lesion (location, histology, T stage, diameter, bleeding, pain, and previous treatment). Continuous covariates were dichotomized and binary variables were normalized to “0” or “1” values. Diameter values were dichotomized as <3 cm and ≥3 cm. The dependent variable was the lesion response to RT (cCR or no cCR). The randomForest package was used to define important explanatory variables. Using the importance() function, we evaluated the importance of each variable. Variables associated with a mean decrease in the Gini coefficient greater than 3% were included in the construction of the classification tree.

Once the most important explanatory variables were identified, the rpart package was employed to build the optimal decision tree, with the option: method=“class”, which generates a classification tree. The tree construction was controlled using the following parameters: complexity parameter (cp)=10^-4 and minbucket (number of observations in any terminal node)=1, to control the tree length prior to pruning (6). A classification tree was then built to optimally partition patients based on cCR. At each node, the algorithm selected the variable and cut-off that produced the greatest separation between internally homogeneous subgroups (7). Pruning was performed using the cross-validation error criterion to define the optimal tree size. The minimum error rule (size producing the minimum cross-validation error) was applied.

Results

A total of 121 patients (65 males) with 145 non-melanoma skin cancer lesions were included in this analysis. The patient, tumor lesion and treatment characteristics are presented in Table I. The mean age at diagnosis was 86.1 years (range=70-100 years) and 28.1% of patients (n=34) were qualified for home health care or resided in a nursing home. Half of the tumor lesions (n=72, 49.7%) had a diameter ≥3 cm. Overall, 79 lesions (54.5%) were painful at the time of diagnosis, whereas bleeding was evident in 56 cases (38.6%). Each lesion received the prescribed hypofractionated RT dose.

Table I. Patient, tumor lesion and treatment characteristics.

T: Tumor; RT: radiotherapy.

The mean follow-up time was 18.3 months (range=3-141 months). Fifty-two patient deaths (43.0%) were recorded. OS rates at 1 year and 2 years were 68.8% [95% confidence interval (CI)=0.585-0.770] and 49.9% (95%CI=0.381-0.606), respectively. Fifty-eight events (recurrence or death) were observed, leading to a 1-year PFS rate of 63.6% (95%CI=0.533-0.722) and a 2-year PFS rate of 43.3% (95%CI=0.314-0.546). Overall, 21 locoregional relapses (14.5%, 21/145 lesions) and 5 distant relapses (4.1%, 5/121 patients) were observed. A cCR was recorded in 114 lesions (78.6%).

To define the significant clinical predictors of cCR, the following variables were investigated: location, histology, T stage, diameter, bleeding, pain, and previous treatment. The dependent variable was the lesion response to RT (cCR or no cCR). The top three important predictors were T stage, location and pain with a mean decrease in the Gini coefficient of 6.32%, 5.86% and 3.24%, respectively (Figure 1). These covariates were used to develop an optimal classification tree (Figure 2). Overall, the tree had five terminal nodes, which resulted in partitioning non-melanoma skin cancer lesions into five groups: (i) T1-2 lesions (67% of the overall sample, with 88% probability of cCR); (ii) T3-4 lesions without pain (25% of the overall sample, with 69% probability of cCR); (iii) painful T3 lesions located in the cheek area or ear, pre/retro-auricular region (3% of the overall, 75% cCR probability); (iv) painful T4 lesions located on the cheek area or ear, pre/retro-auricular region (1% of the overall, 0% cCR probability); (v) painful T3-4 lesions located on the forehead-temples, scalp, nose, orbital/periorbital/chin/neck area or extremities (5% of the overall, 14% cCR probability).

Figure 1.

Important predictor variables. T_: T stage; Site_class: tumor location (forehead-temples or scalp, or nose, or orbital/periorbital/chin/neck area or extremities, cheek area or ear, pre/retro-auricular region); Dim: diameter.

Figure 2.

Classification tree for non-melanoma skin cancer lesions. “Site_class ≥2 yes” indicates lesions located on the forehead-temples or scalp, or nose, or orbital/periorbital/chin/neck area or extremities; “Site_class ≥2 no” indicates lesions located on the cheek area or ear, pre/retro-auricular region. T_ indicates T stage; Pain=1 indicates pain present at diagnosis. The decimal value within each node represents the estimated probability of cCR, and the percentage represents the proportion of lesions in that node.

Two groups had cCR rates below 15% and were classified as unfavorable. The remaining three groups, with higher cCR rates ranging from 69% to 88%, were classified as the favorable group. Subgroup analysis was then performed to explore the survival outcomes based on favorable versus unfavorable groups. One-year OS rates were 71.9% (95%CI= 0.614-0.799) in the favorable group and 28.6% (95%CI=0.04-0.612) in the unfavorable groups (p-value=0.073). One-year PFS rates were 66.0% (95%CI=0.553-0.747) in the favorable group and 28.6% (95%CI=0.04-0.612) in the unfavorable group (p-value=0.110). Because of the very small number of patients/events in the unfavorable group after 12 months, the 2-year survival estimates were not calculated.

Discussion

This multicenter study used machine learning methodologies to analyze the response data of elderly patients with non-melanoma skin cancer treated with definitive hypofractionated RT.

We recorded a cCR rate of 78.6% after treatment. The decision tree approach identified five distinct groups with cCR probabilities of 88%, 69%, 75%, 0% and 14%, respectively. The partition was dependent on T stage, location and pain at diagnosis. The T stagewas the most important factor in the classification tree, highlighting and confirming the importance of developing primary and secondary prevention strategies to ensure early diagnosis (8).

The left side of the tree is dominated by painful lesions with lower probabilities of cCR related to extensive lesions located in the cheek area or ear, pre/retro-auricular region. Interestingly, pain at diagnosis appeared to be an important variable, more commonly selected in randomForest than tumor histology and tumor diameter. This suggests that lesion skin infiltration may have a greater impact on nociceptive superficial somatic pain perception than its superficial extension, positioning pain as a clinical marker of deeper or potentially perineural tumor invasion (9).

Within the current literature on non-melanoma skin cancer, it is challenging to evaluate how variables such as T stage, tumor location, and pain experienced at the time of diagnosis influence the response to treatment. According to the National Comprehensive Cancer Network (NCCN) risk stratification, location (head and neck region) and T stage – considering a size of 2 cm as a discriminant for T classification – are associated with the highest risk for poor outcomes (1, 2). Assuming that cCR can serve as a surrogate for survival outcomes, our results align with the NCCN stratification. The favorable group (groups with cCR probability 69%) was associated with better survival outcomes compared to the unfavorable group (groups with cCR probability <15%), although the difference was not statistically significant. For sure, the modest cohort size reduced the statistical power of the analysis. Data collection from a larger number of patients is ongoing to further describe the outcomes of the patients in these subgroups. Nevertheless, we identified T stage, location and pain at diagnosis as potentially informative predictors of cCR when evaluating definitive hypofractionated RT.

Identifying these groups of patients with non-melanoma skin cancer could have important implications for developing potential treatment strategies. We have presented an intriguing hypothesis-generating result: this decision tree algorithm appears to selectively identify patients less likely to achieve a cCR to RT and therefore those who might benefit most from an intensified treatment. These data generate the hypothesis that combined or sequential systemic therapy could be investigated in patients with painful T4 lesions predicted to have a low probability of cCR, although several important questions remain unanswered (10-12). The impact of age on treatment tolerance and toxicity may reduce its effectiveness, making it essential to weigh the absolute benefit against the risk of additional toxicity before prescribing additional treatment. Another key consideration is the influence of exploratory biomarkers as tumor mutational burden and programmed death-ligand 1 (PD-L1) expression in tumoral tissues are not predictive biomarkers (13) on responsiveness. Our analysis was not powered to support these formal associations.

Study limitations. Its retrospective design, modest sample size, heterogeneous tumor histology and location, and multicenter treatment setting may introduce selection and uncontrolled confounding. Furthermore, 145 lesions arose from 121 patients, and the lesion-level machine-learning analysis did not explicitly account for clustering of multiple lesions within individual patients. Several terminal nodes contained very few lesions, increasing model instability and the risk of overfitting. Although cross-validation was used for tree pruning, the model was developed and evaluated using the same cohort and has not undergone independent external validation. Acute and late toxicity and patient-reported quality of life could not be adequately assessed because of missing retrospective data. Radiotherapy schedules were individualized, potentially introducing additional treatment heterogeneity. Accordingly, the identified associations between T stage, tumor location, pain and cCR should be regarded as exploratory. Prospective multicenter validation in a larger independent cohort, with standardized treatment data, toxicity assessment, and patient-reported outcomes, is required before the decision tree can be used for clinical treatment selection.

Despite these limitations, to our knowledge, this is the first decision tree for elderly patients with non-melanoma skin cancer treated with definitive hypofractionated RT. Reliable pretreatment identification of patients at increased risk of incomplete response could ultimately help guide studies of treatment intensification.

Conclusion

These data showed that T stage, location and pain at diagnosis were of relative importance in predicting cCR in non-melanoma skin cancer treated with definitive hypofractionated RT. Our decision tree could assist clinicians in individualized care of their elderly patients.

Given its potential impact on treatment selection, it is essential that our findings be validated using a prospective multicenter dataset, incorporating patient-reported outcomes analysis.

Conflicts of Interest

The Authors have no conflicts of interest to declare in relation to this study.

Authors’ Contributions

Conceptualization, Francesca De Felice and Anna Merlotti; Data curation, Lavinia Spinelli, Paola Critelli, Melissa Scricciolo, Daniela Musio, Annalisa Magnante and Stefania Martini; Formal analysis, Francesca De Felice; Writing – original draft, Francesca De Felice and Anna Merlotti; Writing – review & editing, Francesca De Felice, Lavinia Spinelli, Paola Critelli, Melissa Scricciolo, Stefania Martini, Giuseppe Minniti and Anna Merlotti.

Funding

This research received no funding.

Artificial Intelligence (AI) Disclosure

No generative artificial intelligence (AI) tools, including large language models or machine learning-based image enhancement tools software, were used in the preparation, analysis, or presentation of this manuscript.