Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2025 Oct 31.
Published in final edited form as: J Natl Cancer Inst. 2025 Nov 1;117(11):2352–2361. doi: 10.1093/jnci/djaf228

Basal cell carcinoma risk prediction in survivors of childhood cancer

Cindy Im 1,†, Christina Boull 1,2, Zhe Lu 3, Kenneth Liao 1, Hasibul Hasan 1, Linwan Xu 3, Yadav Sapkota 4, Rebecca M Howell 5, Michael A Arnold 6, Miriam R Conces 7, Ashley J Housten 8, Judith Gebauer 9, Thorsten Langer 9, Jop C Teepen 10, Leontien C M Kremer 10, Louis S Constine 11, Yutaka Yasui 4, Melissa M Hudson 4,12, Kirsten K Ness 4, Gregory T Armstrong 4,12, Joseph P Neglia 1, Yan Yuan 3,*, Lucie M Turcotte 1,*
PMCID: PMC12574208  NIHMSID: NIHMS2118460  PMID: 40833927

Abstract

Background:

Survivors of childhood cancer face excess risk of developing basal cell carcinoma (BCC). Age-specific BCC risk prediction models for survivors may support targeted screening recommendations.

Methods:

We developed models predicting BCC risk by ages 40 and 50 years featuring detailed cancer treatment predictors, utilizing statistical/machine learning algorithms and data from 23,166 five-year survivors in the Childhood Cancer Survivor Study (CCSS), a multi-institutional retrospective cohort study. Selected models were externally validated in 5,314 survivors in the St. Jude Lifetime Cohort (SJLIFE). Model discrimination and precision were evaluated using the areas under the receiver operating characteristic curve (AUROC) and precision-recall curve (AUPRC), and benchmarked against the current Children’s Oncology Group Long-Term Follow-Up Guidelines (COG LTFU, v6.0) for skin cancer screening.

Results:

By ages 40 and 50 years, BCC cumulative incidence was 5% and 15% in CCSS and 7% and 21% in SJLIFE. The XGBoost algorithm-based models with treatment dose-specific predictors performed best, showing good external discrimination (AUROC40y=0.75; AUROC50y=0.76) and precision (AUPRC40y=0.20; AUPRC50y=0.52), outperforming COG LTFU Guideline-directed risk stratification (AUROC40y=0.65, AUROC50y=0.62; AUPRC40y=0.09, AUPRC50y=0.26; P<0.01). These novel models reclassified 37% of survivors with COG-recommended skin cancer screening as low risk by age 40 and 29% of survivors without COG-recommended screening as moderate/high risk by age 50, suggesting these recommendations overestimate risk in younger survivors and miss relevant predictors (e.g., attained age, chemotherapy).

Conclusions:

In this study, we present validated BCC risk prediction models for childhood cancer survivors that outperform current practice guidelines. The associated online risk calculator can inform risk-/age-based screening recommendations.

INTRODUCTION

With advances in treatment and supportive care, long-term survival after childhood cancer diagnosis now exceeds 85% [1, 2]. However, survivors of childhood cancer face an excess burden of chronic health conditions [3–6], especially subsequent neoplasms (SNs) [7, 8]. Keratinocyte carcinomas (KCs, also known as non-melanoma skin cancer) account for nearly 60% of SNs [9, 10]. Among survivors, basal cell carcinomas (BCCs) are the predominant type, constituting 94% of KCs [11].

Compared to the general population, survivors have a 30-fold greater risk of BCC [9]. Risk factors include radiation therapy (RT), hematopoietic cell transplantation (HCT), and specific chemotherapy drugs (e.g., alkylating agents, anthracyclines, platinum agents, epipodophyllotoxins, vinca alkaloids) [9–13]. RT is the leading risk factor, with robust evidence indicating BCC risk increases with higher therapeutic doses or larger surface areas of irradiation [9–13]. Compared with surgery only, survivors with RT doses ≥35 Gy have a nearly 40-fold greater odds of developing BCCs in the treatment field [9, 12]. HCT is another important risk factor [9, 11, 13], including HCT without total body irradiation (TBI) [11]. While evidence of chemotherapy-related risk is mixed [9, 11, 12], previous work observed survivors treated with RT and chemotherapy had considerably greater excess risk than RT only [9].

Currently, the Children’s Oncology Group Long-Term Follow-Up (COG LTFU) Guidelines [14], the US practice guidelines for childhood cancer survivorship care, recommend that survivors with any history of RT or HCT seek annual dermatologic skin cancer screenings and complete monthly skin self-exams. However, these guidelines do not provide personalized risk assessments or discuss when screenings should begin. Prior studies have found that <30% of at-risk survivors adhere to these recommendations [15–17]. Delayed detection is associated with more invasive surgical treatments and worse outcomes [18–20]. A validated BCC risk stratification tool that informs risks at different time points during follow-up may support screening recommendations and reduce morbidity.

To our knowledge, existing skin cancer risk stratification tools [21–25] do not comprehensively evaluate clinical features relevant to childhood cancer survivors. Using cohort study data including 23,166 long-term survivors, we developed age-specific BCC risk stratification tools that primarily assess treatment predictors followed by an external validation study. Comparison with COG LTFU Guideline-directed skin cancer risk stratification demonstrated potential usability.

PATIENTS AND METHODS

Study population

For model development, we used data from the Childhood Cancer Survivor Study (CCSS), a large, multi-institutional North American retrospective cohort study of five-year survivors diagnosed with pediatric cancer at age <21 years between 1970–1999, with prospective follow-up of outcomes via periodic surveys. CCSS study design and methodology have been described in detail previously [26]. Participants enrolled in the St. Jude Lifetime Cohort (SJLIFE) were excluded in the model development dataset.

Models developed in CCSS were validated using data from SJLIFE, a retrospective cohort study of five-year survivors treated for pediatric cancer between 1962–2012 at St. Jude Children’s Research Hospital, with prospective follow-up of outcomes largely ascertained through clinical assessment. Details regarding SJLIFE design and methodology have been published [27].

Human subjects research approval was granted by the institutional review boards of all participating institutions prior to recruitment. All participants provided informed consent.

BCC ascertainment

In CCSS and SJLIFE, SNs are identified by self-/proxy-report or death certificate review and validated by a pathologist and an oncologist. Participants with complete primary cancer information, unambiguous BCC diagnosis information (cancer site and date), and who had not developed subsequent malignant neoplasms (SMNs) or BCCs within five years of primary cancer diagnosis were included (Figure S1). Only first BCCs with behavior code 3 by the International Classification of Diseases for Oncology [28] (third edition), i.e., SMNs, that developed ≥5 years after childhood cancer diagnosis were included.

Predictors

Relevant predictors, including treatments delivered within five years of diagnosis, were abstracted from medical records. Evaluated predictors included: sex; self-reported race and ethnicity; primary cancer diagnosis and diagnosis age; HCT (yes/no); TBI dose; body region-specific cumulative RT doses directed to the brain, neck, chest, abdomen, pelvis, arms, or legs; and cumulative doses of chemotherapy agents (Table S1). Maximum target radiation doses to body-specific regions were obtained by summing the cumulative delivered dose based on all overlapping fields [29]. Alkylating agents were quantified by the cyclophosphamide-equivalent dose [30], anthracyclines were quantified by the doxorubicin-equivalent dose [31], platinum agents were quantified by the cisplatin-equivalent dose [32, 33], and epipodophyllotoxins were evaluated as cumulative doses of teniposide and etoposide.

Statistical analysis

Figure 1 summarizes the analytic framework to predict age-specific BCC risk, i.e., BCC risks by ages 40 and 50 years. Consistent with our previous work [34], we used an established multiple imputation method based on the predictive mean matching approach [35] to impute missing predictor values in CCSS, generating five imputed datasets. We evaluated three established risk prediction algorithms, considering predictors listed in Table S1: (1) random survival forests [36] (RSF), a tree-based machine learning algorithm for time-to-event data with automatic variable selection; (2) logistic regression (LR), retaining all variables; and (3) XGBoost [37] (eXtreme Gradient Boosting), a tree-based machine learning algorithm that automatically performs variable selection and models non-linear effects and complex interactions. To support the use of binary classification algorithms (LR, XGBoost) while accounting for censoring and competing risk, inverse probability censoring weights [38] (IPCW) were used. Survivors with BCCs occurring before the age threshold of interest were counted as events. Survivors without such events and whose age at last follow-up or death were before the age threshold were censored.

Figure 1:

Figure 1:

Open in a new tab

Subsequent BCC risk prediction model development and validation among survivors of childhood cancer. After applying inclusion/exclusion criteria in age-specific datasets in CCSS, we performed multiple imputation and estimated inverse probability censoring weights (IPCW) to apply binary classification algorithms while appropriately accounting for censoring (with the exception of random survival forests, where imputed datasets with IPCW are unnecessary). Statistical/machine learning algorithms were trained in CCSS and cross-validated prediction performance metrics were evaluated. Trained BCC risk prediction models were then tested in an independent cohort (SJLIFE) and assessed using the same metrics. Potential reclassification of survivors considered to be at risk based on the current Children’s Oncology Group (COG) Long-Term Follow-Up Guidelines was assessed, considering new model-predicted risks stratifying survivors into low, moderate, or high predicted risks for developing BCC by ages 40 and 50 years.

In CCSS, we developed two age-specific models for each statistical/machine learning algorithm: (1) simple models, where all treatment-related predictors were categorized as yes/no exposures and RT was evaluated as a single yes/no predictor; and (2) dose-specific models, with cumulative radiotherapy and chemotherapy dosing (Table S1). Prediction performance was evaluated internally using a cross-validation framework to avoid optimistic estimates of performance. Prediction performance metrics from five-fold cross-validation (LR) or nested cross-validation (XGBoost, RSF) were weighted using IPCW to account for censoring in our data including: the scaled Brier score (SBrS) for overall model performance [39]; area under the receiver operating characteristic curve (AUROC) for model discrimination [40]; and area under the precision-recall curve (AUPRC) for model precision or the average positive predictive value across varying sensitivities [38]. These metrics have a maximum value of one, with larger values indicating better performance relative to metric-specific minimum values (0 for SBrS; 0.5 for AUROC; event rate/incidence for AUPRC). A global calibration statistic, the Spiegelhalter-z statistic [41], was also evaluated, where absolute values closer to zero indicate better calibration. Based on clinician input and previous work [21, 22, 25], we stratified model-predicted risk for subsequent BCC as follows: low (<5%), moderate (≥5% to <20%), and high (≥20%). We assessed the observed BCC cumulative incidence based on model-predicted risk groups in the held-out test data folds from cross-validation procedures in CCSS. Two-sided bootstrapped p-values (1,000 bootstraps) were used to make inferences on performance metrics (AUROCs/AUPRCs) comparing the novel models and COG LTFU Guideline-directed risk stratification. Based on comprehensive assessment of these prediction performance metrics in CCSS, we selected the best models for external validation.

For model external validation in SJLIFE, we computed survivors’ model-predicted risks for developing BCC by ages 40 and 50 years using complete case data, averaging predicted risks from CCSS-trained models across imputed datasets. We evaluated the aforementioned prediction performance metrics, accounting for IPCW, along with calibration plots comparing the observed and model-predicted BCC risks in SJLIFE. The observed BCC cumulative incidence by ages 40 and 50 years in model-predicted risk groups were compared to COG LTFU Guideline-directed risk stratification groups. For each age threshold, we also assessed the novel models’ potential for reclassifying survivors in COG LTFU Guideline-directed risk stratification groups.

Additional methodology details are given in the Supplementary Methods. All statistical analyses were done with R v4.3.0.

RESULTS

A total of 23,166 CCSS survivors (follow-up: median=27.5 years, IQR: 20.3–34.4) were included in the model development dataset; 1,148 survivors (5.0%) developed their first BCC by age 50 years (Figure 1, Table 1). For the external validation study, 5,314 SJLIFE survivors (follow-up: median=18.5 years, IQR: 10.4–27.9) were included and 222 survivors (4.2%) developed their first BCC by age 50 years. Participants were predominantly Non-Hispanic White (CCSS: 75%; SJLIFE: 78%), with similar ages at primary cancer diagnosis across cohorts (CCSS: median=7.0 years; SJLIFE: median=6.6 years). Proportions of survivors treated with TBI or RT were comparable across cohorts. The estimated cumulative incidence of BCC was 5.2% and 14.7% in CCSS and 6.6% and 21.2% in SJLIFE by ages 40 and 50 years, with greater censoring in SJLIFE (Figure S2).

Table 1:

Major characteristics of the CCSS and SJLIFE cohorts

Characteristic CCSS Model Development (N=23166) SJLIFE External Validation (N=5314)
n (%) or median (IQR) n (%) or median (IQR)
Sex
 Female 10750 (46.4%) 2554 (48.1%)
 Male 12416 (53.6%) 2760 (51.9%)
Race and ethnicity
 Hispanic/Latine 1979 (8.5%) 183 (3.4%)
 Non-Hispanic Black 1207 (5.2%) 855 (16.1%)
 Non-Hispanic White 17427 (75.2%) 4148 (78.1%)
 Other 1210 (5.2%) 128 (2.4%)
 Missing 1343 (5.8%) NA
Childhood cancer type
 Leukemia 6846 (29.6%) 1810 (34.1%)
 Central nervous system tumors 4194 (18.1%) 749 (14.1%)
 Hodgkin lymphoma 2756 (11.9%) 536 (10.1%)
 Kidney tumors 2074 (9.0%) 314 (5.9%)
 Non-Hodgkin lymphoma 1905 (8.2%) 333 (6.3%)
 Soft tissue sarcoma 1637 (7.1%) 325 (6.1%)
 Other 3754 (16.2%) 1247 (23.5%)
Age at childhood cancer diagnosis
 Median (IQR) in years 7.0 (3.1 – 13.2) 6.6 (2.8, 12.9)
Age at first BCC or last follow-up
 Median (IQR) in years 40.5 (32.6 – 47.7) 32.3 (23.8, 41.0)
Decade of primary cancer diagnosis
 Before 1980 6122 (26.4%) 737 (13.9%)
 1980–1989 9081 (39.2%) 1205 (22.7%)
 1990–1999 7963 (34.4%) 1444 (27.2%)
 2000–2012 NA 1928 (36.3%)
Hematopoietic cell transplant
 Yes 1186 (5.1%) 441 (8.3%)
 Missing 2078 (9.0%) NA
Total body irradiation
 Yes 536 (2.3%) 147 (2.8%)
 Median (IQR) in cGy 1200 (1121, 1320) 1200 (1200, 1400)
 Missing 776 (3.3%) NA
Cranial RT
 Yes 5500 (23.7%) 1298 (24.4%)
 Median (IQR) in cGy 2400 (1800, 5300) 2600 (2400, 5400)
 Missing 2557 (11.0%) NA
Neck RT
 Yes 3956 (17.1%) 958 (18.0%)
 Median (IQR) in cGy 3400 (2300, 4000) 2600 (2300, 3600)
 Missing 2548 (11.0%) NA
Chest RT
 Yes 4611 (19.9%) 1021 (19.2%)
 Median (IQR) in cGy 3000 (2000, 3900) 2600 (2100, 3500)
 Missing 2559 (11.0%) NA
Abdominal RT
 Yes 4351 (18.8%) 949 (17.9%)
 Median (IQR) in cGy 2600 (2000, 3600) 2340 (1800, 3500)
 Missing 2557 (11.0%) NA
Pelvic RT
 Yes 3560 (15.4%) 841 (15.8%)
 Median (IQR) in cGy 3000 (2000, 3600) 2400 (2000, 3500)
 Missing 2550 (11.0%) NA
Arm RT
 Yes 155 (0.7%) 41 (0.8%)
 Median (IQR) in cGy 4900 (3500, 5500) 4700 (3800, 6000)
 Missing 2543 (11.0%) NA
Leg RT
 Yes 704 (3.0%) 163 (3.1%)
 Median (IQR) in cGy 3600 (2200, 5000) 2600 (2000, 4500)
 Missing 2545 (11.0%) NA
Alkylating agentsa
 Yes 9847 (42.5%) 3038 (57.2%)
 Median (IQR) in mg/m2 7632 (3673, 13470) 7760 (4176, 13007)
 Missing 3540 (15.3%) NA
Anthracyclinesb
 Yes 9119 (39.4%) 2954 (55.6%)
 Median (IQR) in mg/m2 196 (124, 320) 151 (63, 214)
 Missing 2721 (11.7%) NA
Platinumc
 Yes 3916 (16.9%) 855 (16.1%)
 Median (IQR) in mg/m2 467 (308, 650) 466 (303, 782)
 Missing 2142 (9.2%) NA
Epipodophyllotoxinsd
 Yes 3042 (13.2%) 1828 (34.4%)
 Median (IQR) in mg/m2 1699 (900, 3015) 2101 (951, 6017)
 Missing 2301 (9.9%) NA
Glucocorticoids
 Yes 8472 (36.6%) 2227 (41.9%)
 Missing 3023 (13.0%) NA
Vinca alkaloids
 Yes 13663 (59.0%) 3592 (67.6%)
 Missing 3073 (13.3%) NA
Subsequent BCC cumulative incidencee (95% CI)
 By age 40 years 5.2 (4.9–5.6) 6.6 (5.5–7.8)
 By age 50 years 14.7 (14.0–15.6) 21.2 (18.4–24.0)
Open in a new tab

Abbreviations: cGy, centigray; mg, milligrams; m, meters; IQR, interquartile range; CI, 95% confidence interval. Median dose and IQR provided for survivors who received the corresponding therapy. Other race and ethnicity includes Asian/Pacific Islander, American Indian/Alaskan Native, and multiple race survivors.

a.

Provided as cyclophosphamide-equivalent dose.

b.

Provided as doxorubicin-equivalent dose.

c.

Provided as cisplatin-equivalent dose.

d.

Includes doses of etoposide and teniposide.

e.

Estimated accounting for inverse probability censoring weights.

Prediction performance in CCSS

In CCSS, SBrS metrics for overall performance indicated the dose-specific models performed similarly to the simple models in predicting BCC risks by age 40 (e.g., dose-specific SBrSXGBoost=4.5%, simple SBrSXGBoost=4.2%) but was modestly better for prediction by age 50 (dose-specific: SBrSXGBoost=14.5%, simple: SBrSXGBoost=11.8%, Table S2). The XGBoost-based models showed evidence of having the best overall performance among algorithms at both age thresholds, especially with respect to precision (Figures S3–S4). The XGBoost-based dose-specific model performed best for predicting BCC by age 50, with good discrimination (AUROCXGBoost=0.75) and precision (AUPRCXGBoost=0.42). Global calibration statistics and the observed BCC cumulative incidences in model-predicted risk groups suggested the XGBoost models were well-calibrated (Table S3). All novel dose-specific and simple models had significantly better discrimination (P<0.01) and precision (P<0.01) compared with COG LTFU Guideline-directed risk stratification (age 40: AUROCCOG=0.64, AUPRCCOG=0.07; age 50: AUROCCOG=0.61, AUPRCCOG=0.18).

Shared top XGBoost predictors (i.e., median ≥5% variable importance gain across imputed datasets) between dose-specific and simple models included age at childhood cancer diagnosis, HCT, and RT (Figure 2, Figure S5). Age at diagnosis and RT (particularly head/neck dose) showed the largest variable importance gains (>10%). While anthracycline and alkylator dose were also among top predictors for the dose-specific model, childhood cancer diagnosis type appeared to compensate for the loss of chemotherapy dose information in the simple model. Further assessment of the dose-specific and simple XGBoost models revealed clear BCC risk differentiation between the model-predicted risk groups (Table S4). For example, odds ratios (ORs) comparing moderate and high versus low XGBoost-predicted risk groups from the dose-specific model were distinct at age 40 (ORmoderate=4.12, 95% CI: 3.54–4.80; ORhigh=13.81, 95% CI: 10.44–18.27) and age 50 (ORmoderate=2.73, 95% CI: 2.17–3.43; ORhigh=10.59, 95% CI: 8.44–13.29).

Figure 2:

Figure 2:

Open in a new tab

Influential predictors of BCC risk by age 50 years in CCSS. The distribution of the XGBoost model variable importance gain in the imputed datasets for predictors that were among the top predictors with a median gain of at least 5% in all five imputed datasets for the dose-specific model. Abbreviations: CC DX, childhood cancer diagnosis; HCT, hematopoietic cell transplantation; RT, radiation therapy.

External validation in SJLIFE

We selected the simple and dose-specific XGBoost-based models for validation. In SJLIFE, these models had similarly good discrimination (age 40: AUROCsimple=0.74, AUROCdose-specific=0.75; age 50: AUROCsimple=0.76, AUROCdose-specific=0.76) and precision (age 40: AUPRCsimple=0.20, AUPRCdose-specific=0.20; age 50: AUPRCsimple=0.51, AUPRCdose-specific=0.52) (Figure 3, Table S5). COG LTFU Guideline-directed risk stratification for skin cancer remained significantly less predictive (age 40: AUROCCOG=0.65, AUPRCCOG=0.09, P<0.01; age 50: AUROCCOG=0.62, AUPRCCOG=0.26, P<0.01).

Figure 3:

Figure 3:

Open in a new tab

Subsequent BCC risk prediction model performance in the SJLIFE model validation cohort. Top panels depict receiver operating characteristic (ROC) curves for XGBoost models with dose-specific (blue) and simple (black) predictors at different age thresholds while bottom panels show corresponding precision-recall (PR) curves. Red curves represent ROC and PR curves using a classifier based on the Children’s Oncology Group (COG) Long-Term Follow-Up Guidelines for recommending skin cancer screening. Bootstrapped two-sided p-values comparing differences in prediction performance metrics from novel models compared to COG Guideline-directed risk stratification are provided.

Calibration across age thresholds for dose-specific and simple models were similar and reasonable in SJLIFE (Figure S6), but the dose-specific model showed modestly better calibration and risk differentiation by age 50 (Tables S6–S8), i.e., identifying fewer survivors at moderate-risk (dose-specific: 38.4%; simple: 43.7%) and more distinction in odds by risk group (dose-specific: ORmoderate=12.05, 95% CI: 5.43–26.72; ORhigh=31.62, 95% CI: 14.39–69.45 versus simple: ORmoderate=7.15, 95% CI: 3.61–14.20; ORhigh=19.73, 95% CI: 10.01–38.89). All novel models had superior risk stratification compared with the COG LTFU Guidelines and confirmed three distinct risk groups (Figure 4, Figure S7). For example, the observed BCC cumulative incidence by age 40 was 9.5% among survivors classified as at-risk by the COG LTFU Guidelines. In contrast, for the dose-specific model, the observed cumulative incidence was 12.4% (95% CI: 10.4–14.5%) in the moderate-risk group and 26.5% (95% CI: 13.9–40.3%) in the high-risk group. By age 50, the cumulative incidence in the moderate-risk group (dose-specific: 18.2%, 95% CI: 15.1–21.6%) remained differentiated from the high-risk group (dose-specific: 36.9%, 95% CI: 33.0–41.1%), where the latter exceeded the cumulative incidence in survivors at-risk based on the COG LTFU Guidelines (26.4%). Among survivors not considered to be at-risk per COG LTFU Guidelines, the cumulative incidence of BCC surpassed those with low model-predicted risk (COG: 6.2%; dose-specific: 1.8%, 95% CI: 0.8–3.9%).

Figure 4:

Figure 4:

Open in a new tab

BCC cumulative incidence by specified age thresholds in dose-specific model-predicted risk groups in the independent SJLIFE validation cohort and reclassification of Children’s Oncology Group (COG) Long-Term Follow-Up Guideline-directed risk groups. In panel A, IPCW-based BCC cumulative incidence is provided. Risk groups are defined by the novel XGBoost risk prediction models (left) and the Children’s Oncology Group Long-Term Follow-Up Guidelines (right). In panel B, pie charts categorizing survivors into low (<5%; blue), moderate (5–19%; orange), and high (≥20%; red) risk groups using model-predicted risks from the XGBoost risk prediction models for different age thresholds are shown, separating survivors recommended screening per COG (top row) from those who are not recommended screening (bottom row).

Upon investigating reclassification of survivors’ risk designations informed by the COG LTFU Guidelines, we found 36.8% of SJLIFE survivors who would be considered at-risk following the COG LTFU Guidelines would be reclassified as low-risk by age 40 using the novel dose-specific XGBoost-based model (Figure 4, Table S8). By age 50, this proportion was 8.1%. Notably, 29.2% of survivors who would not be considered to be at-risk according to the COG LTFU Guidelines were identified as having moderate- or high-predicted risk by age 50 using this model.

DISCUSSION

To our knowledge, the current study provides the first BCC risk prediction models available for clinical use among survivors of childhood cancer and their care providers. These models were developed in a large cohort of survivors with pathology-ascertained BCCs and leveraged an age-specific machine learning (XGBoost) algorithm. In external validation data, our XGBoost-based treatment dose-specific and simple models showed good discrimination and precision across different assessment ages. Importantly, these novel models: (a) demonstrated superior discrimination and precision compared with BCC risk stratification following COG LTFU Guidelines (P<0.01); (b) placed survivors into three distinct risk categories; and (c) reclassified COG LTFU Guidelines-directed risks for substantial proportions of survivors. These novel risk stratification tools are available as easy-to-use online risk calculators (ccss.stjude.org/basalcalc). In general, our results suggest users should expect dose-specific and simple models to provide similar age-specific BCC risk estimates on average, but the dose-specific model more efficiently identifies survivors at low- or high-risk. Detailed examples describing how tools may be applied are provided in the Supplementary Materials.

By 2040, an estimated 580,000 childhood cancer survivors will be living in the United States alone [42]. Individualized late effects risk assessment tools that can improve quality of life in this growing survivor population are needed. Currently, annual skin cancer screening via full-body skin examination after treatment with RT or HCT is recommended by the COG LTFU Guidelines [14]. However, <30% of at-risk survivors are adherent [15–17]. In alignment with existing interventions to improve screening rates [16], our novel BCC risk stratification tools may facilitate targeted efforts to educate survivors based on their risk categorizations at ages 40 and 50 years. Given that our novel models identify three distinct risk groups at both age thresholds, the results of this study could inform future studies evaluating tailored risk-based referral and screening strategies, e.g., adjusting the intensity of counseling, timing of first dermatologic referrals, and prescribed frequency of full-body skin exams, depending on the ages at which survivors may have moderate or high model-predicted risks.

To date, the COG LTFU Guidelines provide no recommendations for when skin cancer screenings should begin for at-risk survivors. This lack of specificity has concerning implications. First, despite the fact that BCCs can be successfully treated and are rarely fatal, timely dermatologic evaluation and treatment is necessary to avoid excess morbidity [18–20, 43]. Childhood cancer survivors face greater risk of financial hardship [44, 45] and may also have higher patient time costs during follow-up [46]. Therefore, some survivors may not have timely access to dermatologic care. Second, promoting full-body skin exams in an unselected population is not cost-effective [24, 47]. Age-based screening recommendations may be key for improving adherence and supporting the judicious use of dermatology resources. The observed BCC cumulative incidence by age 40 years among survivors considered to be at-risk following the COG LTFU Guidelines is 7.5–9.5%, potentially placing an excess burden of unnecessary dermatologic screening on young adult survivors. Compared to existing guidelines, our novel models can identify many survivors who could be advised of their lower screening burden before age 40 while also identifying survivors without RT or HCT who could be informed of their moderate-to-high predicted risk of developing BCC by age 50.

Several limitations should be considered in interpreting these results. Despite the strengths of using a large cohort of survivors for model training, racial/ethnic diversity in CCSS is limited. Furthermore, the lack of external validation studies in survivors outside the US is a limitation. Future studies in international survivor cohorts [9] are needed. Modifiable cumulative lifetime exposures, including UV exposure and lifestyle factors (e.g., smoking, sun protection), were not included given these models were intentionally anchored to evaluate exposures at the five-year cancer survival mark. While previous work suggests these predictors are less potent risk factors in survivors [11, 12], other predictors identified in the general population skin cancer literature that are assessable at five-year survival including skin phototype and genetics [23–25] could improve predictive performance, especially in survivors not treated with RT. Unfortunately, data related to these predictors were not available for most (>60%) CCSS participants and reported race and ethnicity was used as a less specific marker of UV susceptibility (versus skin phototype). Observations of substantially lower rates of BCC among irradiated Black survivors than expected [48] suggest these additional host photosensitivity characteristics should be assessed in future studies.

CCSS also does not systematically capture data related to treatment beyond the five-year period post-diagnosis, e.g., the treatment of SMNs, or skin RT dose-volume and surface area; incorporating such information could potentially improve risk prediction performance. Models to predict risks of developing BCCs in the radiation field were not developed given previous work reporting >85% of BCCs identified in CCSS developed in the RT field [11] and the lower statistical power for developing similar models using data limited to irradiated survivors. Lastly, these models were developed in study data with survivors diagnosed between 1970–1999. Therefore, these do not account for more contemporary treatment predictors, e.g., immunotherapy or targeted therapies. Although these models will require updates when such data become available, these are less relevant in supporting the current need for BCC clinical risk stratification among the growing number of survivors who are now approaching their fourth and fifth decades of life.

In summary, we developed and validated age-specific BCC risk stratification tools for survivors of childhood cancer. These novel models were significantly more predictive than COG LTFU Guideline-directed risk stratification and provided clinically relevant risk reclassification. These novel models can inform age-based skin cancer screening recommendations, which are currently unavailable for this high-risk population.

Supplementary Material

40833927 Supplemental

Acknowledgements

The funding sources had no role in the study design, the collection, analysis, or interpretation of data, or writing. The corresponding author had full access to data used to develop and validate models and had final responsibility for the decision to submit for publication. This study was presented in part at the International Symposium on Late Complications after Childhood Cancer (June 28, 2024) and the 56th Annual Congress of the International Society of Paediatric Oncology (October 20, 2024).

Funding

This work was funded by the US National Cancer Institute (R21 CA261833, C Im/Y Yuan principal investigators; R01 CA216354, Y Yasui/J Zhang, MPIs), Canadian Institutes of Health Research (FRN 148693, Y Yuan principal investigator), and the Children’s Cancer Research Fund (C Im, C Boull, LM Turcotte). C. Im is also supported by the US National Heart, Blood, and Lung Institute (R01 HL173881, C Im/Y Sapkota, principal investigators) and the University of Minnesota Foundation Pediatric Scholar Award. The study data was from the Childhood Cancer Survivor Study (U24 CA55727, GT Armstrong principal investigator) and the St. Jude Lifetime Cohort (U01 CA195547, MM Hudson/KK Ness, principal investigators). Support to St. Jude Children’s Research Hospital was also provided by the Cancer Center Support (CORE) grant (CA21765, C. Roberts, Principal Investigator) and the American Lebanese-Syrian Associated Charities (ALSAC).

Footnotes

Conflicts of interest

The authors declare no competing interests.

Data availability statement

The Childhood Cancer Survivor Study is a US National Cancer Institute funded resource (U24 CA55727) to promote and facilitate research among long-term survivors of cancer diagnosed during childhood and adolescence. CCSS data are publicly available on dbGaP at https://www.ncbi.nlm.nih.gov/gap/ through its accession number phs001327.v2.p1 and on the St Jude Survivorship Portal within the St. Jude Cloud at https://survivorship.stjude.cloud/. In addition, utilization of the CCSS data that leverages the expertise of CCSS Statistical and Survivorship research and resources will be considered on a case-by case basis. For this utilization, a research Application Of Intent followed by an Analysis Concept Proposal must be submitted for evaluation by the CCSS Publications Committee. Users interested in utilizing this resource are encouraged to visit http://ccss.stjude.org. Full analytical data sets associated with CCSS publications since January of 2023 are also available on the St. Jude Survivorship Portal at https://viz.stjude.cloud/community/cancer-survivorship-community~4/publications. St. Jude Lifetime Cohort (SJLIFE) data used for model validation may be accessed from the St. Jude Cloud (https://stjude.cloud) under the accession number SJC-DS-1002.

REFERENCES

  • 1.Howlader N, Noone A, Krapcho M, et al. SEER Cancer Statistics Review, 1975–2016. Bethesda, MD: National Cancer Institute. In; 2019. [Google Scholar]
  • 2.Armstrong GT, Chen Y, Yasui Y, et al. Reduction in late mortality among 5-year survivors of childhood cancer. New England Journal of Medicine 2016;374(9):833–842. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 3.Bhakta N, Liu Q, Ness KK, et al. The cumulative burden of surviving childhood cancer: an initial report from the St Jude Lifetime Cohort Study (SJLIFE). Lancet 2017;390(10112):2569–2582. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 4.Robison LL, Hudson MM. Survivors of childhood and adolescent cancer: life-long risks and responsibilities. Nature Reviews: Cancer 2014;14(1):61. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5.Oeffinger KC, Mertens AC, Sklar CA, et al. Chronic health conditions in adult survivors of childhood cancer. N Engl J Med 2006;355(15):1572–82. [DOI] [PubMed] [Google Scholar]
  • 6.Hudson MM, Ness KK, Gurney JG, et al. Clinical ascertainment of health outcomes among adults treated for childhood cancer. Jama 2013;309(22):2371–2381. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7.Turcotte LM, Liu Q, Yasui Y, et al. Temporal trends in treatment and subsequent neoplasm risk among 5-year survivors of childhood cancer, 1970–2015. Jama 2017;317(8):814–824. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8.Teepen JC, Kremer L, Ronckers CM, et al. Long-term risk of subsequent malignant neoplasms after treatment of childhood cancer in the DCOG LATER study cohort: role of chemotherapy. Journal of Clinical Oncology 2017;35(20):2288–2298. [DOI] [PubMed] [Google Scholar]
  • 9.Teepen JC, Kok JL, Kremer LC, et al. Long-term risk of skin cancer among childhood cancer survivors: a DCOG-LATER cohort study. JNCI: Journal of the National Cancer Institute 2019;111(8):845–853. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10.Perkins JL, Liu Y, Mitby PA, et al. Nonmelanoma skin cancer in survivors of childhood and adolescent cancer: a report from the childhood cancer survivor study. J Clin Oncol 2005;23(16):3733–3741. [DOI] [PubMed] [Google Scholar]
  • 11.Boull C, Chen Y, Im C, et al. Keratinocyte carcinomas in survivors of childhood cancer: A report from the childhood cancer survivor study. Journal of the American Academy of Dermatology 2024. [Google Scholar]
  • 12.Watt TC, Inskip PD, Stratton K, et al. Radiation-related risk of basal cell carcinoma: a report from the Childhood Cancer Survivor Study. Journal of the National Cancer Institute 2012;104(16):1240–1250. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 13.Schwartz JL, Kopecky KJ, Mathes RW, et al. Basal cell skin cancer after total-body irradiation and hematopoietic cell transplantation. Radiation research 2009;171(2):155–163. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14.Children’s Oncology Group. Long-Term Follow-Up Guidelines for Survivors of Childhood, Adolescent and Young Adult Cancers, Version 6.0. www.survivorshipguidelines.org.
  • 15.Geller AC, Keske RR, Haneuse S, et al. Skin cancer early detection practices among adult survivors of childhood cancer treated with radiation. Journal of Investigative Dermatology 2019;139(9):1898–1905. e2. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 16.Geller AC, Coroiu A, Keske RR, et al. Advancing Survivors Knowledge (ASK Study) of skin cancer surveillance after childhood cancer: a randomized controlled trial in the Childhood Cancer Survivor Study. Journal of Clinical Oncology 2023;41(12):2269–2280. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17.Yan AP, Chen Y, Henderson TO, et al. Adherence to surveillance for second malignant neoplasms and cardiac dysfunction in childhood cancer survivors: a childhood cancer survivor study. Journal of Clinical Oncology 2020;38(15):1711–1722. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18.Kricker A, Armstrong B, Hansen V, et al. Basal cell carcinoma and squamous cell carcinoma growth rates and determinants of size in community patients. Journal of the American Academy of Dermatology 2014;70(3):456–464. [DOI] [PubMed] [Google Scholar]
  • 19.Weinstock MA, Still JM. Assessing current treatment options for patients with severe/advanced basal cell carcinoma. In: Seminars in cutaneous medicine and surgery, 2011: Abstract 30, p. S10–3. [DOI] [PubMed] [Google Scholar]
  • 20.Karimkhani C, Boyers LN, Dellavalle RP, et al. It’s time for “keratinocyte carcinoma” to replace the term “nonmelanoma skin cancer”. Journal of the American Academy of Dermatology 2015;72(1):186–187. [DOI] [PubMed] [Google Scholar]
  • 21.Gómez-Tomás Á, Bavinck JNB, Genders R, et al. External validation of the Skin and UV Neoplasia Transplant Risk Assessment Calculator (SUNTRAC) in a large European solid organ transplant recipient cohort. JAMA dermatology 2023;159(1):29–36. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 22.Jambusaria-Pahlajani A, Crow LD, Lowenstein S, et al. Predicting skin cancer in organ transplant recipients: development of the SUNTRAC screening tool using data from a multicenter cohort study. Transplant International 2019;32(12):1259–1267. [DOI] [PubMed] [Google Scholar]
  • 23.Fontanillas P, Alipanahi B, Furlotte NA, et al. Disease risk scores for skin cancers. Nature communications 2021;12(1):160. [Google Scholar]
  • 24.Nagarajan P, Asgari MM, Green AC, et al. Keratinocyte carcinomas: current concepts and future research priorities. Clinical Cancer Research 2019;25(8):2379–2391. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 25.Whiteman DC, Thompson BS, Thrift AP, et al. A model to predict the risk of keratinocyte carcinomas. Journal of Investigative Dermatology 2016;136(6):1247–1254. [DOI] [PubMed] [Google Scholar]
  • 26.Robison LL, Armstrong GT, Boice JD, et al. The Childhood Cancer Survivor Study: a National Cancer Institute–supported resource for outcome and intervention research. Journal of clinical oncology 2009;27(14):2308. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 27.Howell CR, Bjornard KL, Ness KK, et al. Cohort profile: the St. Jude Lifetime Cohort Study (SJLIFE) for paediatric cancer survivors. International Journal of Epidemiology 2021;50(1):39–49. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 28.Jack A, Percy CL, Sobin L, et al. International classification of diseases for oncology: ICD-O: World Health Organization; 2000. [Google Scholar]
  • 29.Howell RM, Smith SA, Weathers RE, et al. Adaptations to a Generalized Radiation Dose Reconstruction Methodology for Use in Epidemiologic Studies: An Update from the MD Anderson Late Effect Group. Radiation research 2019;192(2):169–188. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 30.Green DM, Nolan VG, Goodman PJ, et al. The cyclophosphamide equivalent dose as an approach for quantifying alkylating agent exposure: a report from the Childhood Cancer Survivor Study. Pediatric blood & cancer 2014;61(1):53–67. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 31.Feijen EA, Leisenring WM, Stratton KL, et al. Derivation of anthracycline and anthraquinone equivalence ratios to doxorubicin for late-onset cardiotoxicity. JAMA oncology 2019;5(6):864–871. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 32.Taylor A, Wiltshaw E, Gore M, et al. Long-term follow-up of the first randomized study of cisplatin versus carboplatin for advanced epithelial ovarian cancer. Journal of clinical oncology 1994;12(10):2066–2070. [DOI] [PubMed] [Google Scholar]
  • 33.Mangioni C, Bolis G, Pecorelli S, et al. Randomized trial in advanced ovarian cancer comparing cisplatin and carboplatin. JNCI: Journal of the National Cancer Institute 1989;81(19):1464–1471. [DOI] [PubMed] [Google Scholar]
  • 34.Im C, Lu Z, Mostoufi-Moab S, et al. Development and validation of age-specific risk prediction models for primary ovarian insufficiency in long-term survivors of childhood cancer: a report from the Childhood Cancer Survivor Study and St Jude Lifetime Cohort. The Lancet Oncology 2023;24(12):1434–1442. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 35.Van Buuren S Flexible imputation of missing data: CRC press; 2018. [Google Scholar]
  • 36.Ishwaran H, Kogalur UB, Blackstone EH, et al. Random survival forests. 2008. [Google Scholar]
  • 37.Chen T, Guestrin C. Xgboost: A scalable tree boosting system. In: Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining, 2016, p. 785–794. [Google Scholar]
  • 38.Yuan Y, Zhou QM, Li B, et al. A threshold-free summary index of prediction accuracy for censored time to event data. Statistics in medicine 2018;37(10):1671–1681. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 39.Hu B, Palta M, Shao J. Properties of R2 statistics for logistic regression. Statistics in medicine 2006;25(8):1383–1395. [DOI] [PubMed] [Google Scholar]
  • 40.Pepe MS. The statistical evaluation of medical tests for classification and prediction: Oxford University Press, USA; 2003. [Google Scholar]
  • 41.Spiegelhalter DJ. Probabilistic prediction in patient management and clinical trials. Statistics in medicine 1986;5(5):421–433. [DOI] [PubMed] [Google Scholar]
  • 42.Ehrhardt MJ, Krull KR, Bhakta N, et al. Improving quality and quantity of life for childhood cancer survivors globally in the twenty-first century. Nature Reviews Clinical Oncology 2023;20(10):678–696. [Google Scholar]
  • 43.Wetzel M, Jung JY, Brown TS. Multiple Lesions in Irradiated Skin. JAMA oncology 2019;5(5):728–729. [DOI] [PubMed] [Google Scholar]
  • 44.Nathan PC, Huang I-C, Chen Y, et al. Financial hardship in adult survivors of childhood cancer in the era after implementation of the Affordable Care Act: a report from the Childhood Cancer Survivor Study. Journal of Clinical Oncology 2023;41(5):1000–1010. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 45.Nathan PC, Henderson TO, Kirchhoff AC, et al. Financial hardship and the economic effect of childhood cancer survivorship. Journal of Clinical Oncology 2018;36(21):2198–2205. [DOI] [PubMed] [Google Scholar]
  • 46.Yabroff KR, Guy GP Jr, Ekwueme DU, et al. Annual patient time costs associated with medical care among cancer survivors in the United States. Medical care 2014;52(7):594–601. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 47.Bibbins-Domingo K, Grossman DC, Curry SJ, et al. Screening for skin cancer: US Preventive Services Task Force recommendation statement. Jama 2016;316(4):429–435. [DOI] [PubMed] [Google Scholar]
  • 48.Ehrhardt MJ, Bhakta N, Liu Q, et al. Absence of basal cell carcinoma in irradiated childhood cancer survivors of black race: a report from the St. Jude Lifetime Cohort Study. Cancer Epidemiology, Biomarkers & Prevention 2016;25(9):1356–1360. [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

40833927 Supplemental
NIHMS2118460-supplement-40833927_Supplemental.pdf (1.8MB, pdf)

Data Availability Statement

The Childhood Cancer Survivor Study is a US National Cancer Institute funded resource (U24 CA55727) to promote and facilitate research among long-term survivors of cancer diagnosed during childhood and adolescence. CCSS data are publicly available on dbGaP at https://www.ncbi.nlm.nih.gov/gap/ through its accession number phs001327.v2.p1 and on the St Jude Survivorship Portal within the St. Jude Cloud at https://survivorship.stjude.cloud/. In addition, utilization of the CCSS data that leverages the expertise of CCSS Statistical and Survivorship research and resources will be considered on a case-by case basis. For this utilization, a research Application Of Intent followed by an Analysis Concept Proposal must be submitted for evaluation by the CCSS Publications Committee. Users interested in utilizing this resource are encouraged to visit http://ccss.stjude.org. Full analytical data sets associated with CCSS publications since January of 2023 are also available on the St. Jude Survivorship Portal at https://viz.stjude.cloud/community/cancer-survivorship-community~4/publications. St. Jude Lifetime Cohort (SJLIFE) data used for model validation may be accessed from the St. Jude Cloud (https://stjude.cloud) under the accession number SJC-DS-1002.

RESOURCES