Incorporating spatial dose metrics in machine learning-based normal tissue complication probability (NTCP) models of severe acute dysphagia resulting from head and neck radiotherapy

Risk prediction models for dysphagia after radiotherapy among patients with head and neck cancer: a systematic review and meta-analysis

Background

Predictive models can identify patients at risk and thus enable personalized interventions. Despite the increasing number of prediction models used to predict the risk of dysphagia after radiotherapy in patients with head and neck cancer (HNC), there is still uncertainty about the effectiveness of these models in clinical practice and about the quality and applicability of future studies. The aim of this study was to systematically evaluate and analyze all predictive models used to predict dysphagia in patients with HNC after radiotherapy.

Methods

PubMed, Cochrane Library, EMbase and Web of Science databases were searched from database establishment to August 31, 2024. Data from selected studies were extracted using predefined tables and the quality of the predictive modelling studies was assessed using the PROBAST tool. Meta-analysis of the predictive performance of the model was performed using the "metafor" package in R software.

Results

Twenty-five models predicting the risk of dysphagia after radiotherapy in patients with HNC were included, covering a total of 8,024 patients. Common predictors include mean dose to pharyngeal constrictor muscles, treatment setting, and tumor site. Of these models, most were constructed based on logistic regression, while only two studies used machine learning methods. The area under the receiver operating characteristic curve (AUC) reported values for these models ranged from 0.57 to 0.909, with 13 studies having a combined AUC value of 0.78 (95% CI: 0.74-0.81). All studies showed a high risk of bias as assessed by the PROBAST tool.

Conclusion

Most of the published prediction models in this study have good discrimination. However, all studies were considered to have a high risk of bias based on PROBAST assessments. Future studies should focus on large sample size and rigorously designed multicenter external validation to improve the reliability and clinical applicability of prediction models for dysphagia after radiotherapy for HNC.

Systematic review registration

https://www.crd.york.ac.uk/prospero, identifier CRD42024587252.

A multi-center, multi-organ, multi-omic prediction model for treatment-induced severe oral mucositis in nasopharyngeal carcinoma

PURPOSE

Oral mucositis (OM) is one of the most prevalent and crippling treatment-related toxicities experienced by nasopharyngeal carcinoma (NPC) patients receiving radiotherapy (RT), posing a tremendous adverse impact on quality of life. This multi-center study aimed to develop and externally validate a multi-omic prediction model for severe OM.

METHODS

Four hundred and sixty-four histologically confirmed NPC patients were retrospectively recruited from two public hospitals in Hong Kong. Model development was conducted on one institution (n = 363), and the other was reserved for external validation (n = 101). Severe OM was defined as the occurrence of CTCAE grade 3 or higher OM during RT. Two predictive models were constructed: 1) conventional clinical and DVH features and 2) a multi-omic approach including clinical, radiomic and dosiomic features.

RESULTS

The multi-omic model, consisting of chemotherapy status and radiomic and dosiomic features, outperformed the conventional model in internal and external validation, achieving AUC scores of 0.67 [95% CI: (0.61, 0.73)] and 0.65 [95% CI: (0.53, 0.77)], respectively, compared to the conventional model with 0.63 [95% CI: (0.56, 0.69)] and 0.56 [95% CI: (0.44, 0.67)], respectively. In multivariate analysis, only the multi-omic model signature was significantly correlated with severe OM in external validation (p = 0.017), demonstrating the independent predictive value of the multi-omic approach.

CONCLUSION

A multi-omic model with combined clinical, radiomic and dosiomic features achieved superior pre-treatment prediction of severe OM. Further exploration is warranted to facilitate improved clinical decision-making and enable more effective and personalized care for the prevention and management of OM in NPC patients.

Predictive Factors for Chemoradiation-Induced Oral Mucositis and Dysphagia in Head and Neck Cancer: A Scoping Review

Despite advances in head and neck cancer treatment, virtually all patients experience chemoradiation-induced toxicities. Oral mucositis (OM) and dysphagia are among the most prevalent and have a systemic impact on patients, hampering treatment outcome and harming quality of life. Accurate prediction of severe cases is crucial for improving management strategies and, ultimately, patient outcomes. This scoping review comprehensively maps the reported predictors and critically evaluates the performance, methodology, and reporting of predictive models for these conditions. A total of 174 studies were identified from database searches, with 73 reporting OM predictors, 97 reporting dysphagia predictors, and 4 reporting both OM and dysphagia predictors. These predictors included patient demographics, tumor classification, chemoradiotherapy regimen, radiation dose to organs-at-risk, genetic factors, and results of clinical laboratory tests. Notably, many studies only conducted univariate analysis or focused exclusively on certain predictor types. Among the included studies, numerous predictive models were reported: eight for acute OM, five for acute dysphagia, and nine for late dysphagia. The area under the receiver operating characteristic curve (AUC) ranged between 0.65 and 0.81, 0.60 and 0.82, and 0.70 and 0.85 for acute oral mucositis, acute dysphagia, and late dysphagia predictive models, respectively. Several areas for improvement were identified, including the need for external validation with sufficiently large sample sizes, further standardization of predictor and outcome definitions, and more comprehensive reporting to facilitate reproducibility.

Artificial intelligence-assisted selection and efficacy prediction of antineoplastic strategies for precision cancer therapy

The rapid development of artificial intelligence (AI) technologies in the context of the vast amount of collectable data obtained from high-throughput sequencing has led to an unprecedented understanding of cancer and accelerated the advent of a new era of clinical oncology with a tone of precision treatment and personalized medicine. However, the gains achieved by a variety of AI models in clinical oncology practice are far from what one would expect, and in particular, there are still many uncertainties in the selection of clinical treatment options that pose significant challenges to the application of AI in clinical oncology. In this review, we summarize emerging approaches, relevant datasets and open-source software of AI and show how to integrate them to address problems from clinical oncology and cancer research. We focus on the principles and procedures for identifying different antitumor strategies with the assistance of AI, including targeted cancer therapy, conventional cancer therapy, and cancer immunotherapy. In addition, we also highlight the current challenges and directions of AI in clinical oncology translation. Overall, we hope this article will provide researchers and clinicians with a deeper understanding of the role and implications of AI in precision cancer therapy, and help AI move more quickly into accepted cancer guidelines.

Machine learning for the prediction of toxicities from head and neck cancer treatment: A systematic review with meta-analysis

INTRODUCTION

The aim of the present systematic review (SR) is to summarize Machine Learning (ML) models currently used to predict head and neck cancer (HNC) treatment-related toxicities, and to understand the impact of image biomarkers (IBMs) in prediction models (PMs). The present SR was conducted following the guidelines of the PRISMA 2022 and registered in PROSPERO database (CRD42020219304).

METHODS

The acronym PICOS was used to develop the focused review question (Can PMs accurately predict HNC treatment toxicities?) and the eligibility criteria. The inclusion criteria enrolled Prediction Model Studies (PMSs) with patient cohorts that were treated for HNC and developed toxicities. Electronic database search encompassed PubMed, EMBASE, Scopus, Cochrane Library, Web of Science, LILACS, and Gray Literature (Google Scholar and ProQuest). Risk of Bias (RoB) was assessed through PROBAST and the results were synthesized based on the data format (with and without IBMs) to allow comparison.

RESULTS

A total of 28 studies and 4,713 patients were included. Xerostomia was the most frequently investigated toxicity (17; 60.71 %). Sixteen (57.14 %) studies reported using radiomics features in combination with clinical or dosimetrics/dosiomics for modelling. High RoB was identified in 23 studies. Meta-analysis (MA) showed an area under the receiver operating characteristics curve (AUROC) of 0.82 for models with IBMs and 0.81 for models without IBMs (p value < 0.001), demonstrating no difference among IBM- and non-IBM-based models.

DISCUSSION

The development of a PM based on sample-specific features represents patient selection bias and may affect a model's performance. Heterogeneity of the studies as well as non-standardized metrics prevent proper comparison of studies, and the absence of an independent/external test does not allow the evaluation of the model's generalization ability.

CONCLUSION

IBM-featured PMs are not superior to PMs based on non-IBM predictors. The evidence was appraised as of low certainty.

Cluster-Based Toxicity Estimation of Osteoradionecrosis via Unsupervised Machine Learning: Moving Beyond Single Dose-Parameter Normal Tissue Complication Probability by Using Whole Dose-Volume Histograms for Cohort Risk Stratification

Purpose

Given the limitations of extant models for normal tissue complication probability estimation for osteoradionecrosis (ORN) of the mandible, the purpose of this study was to enrich statistical inference by exploiting structural properties of data and provide a clinically reliable model for ORN risk evaluation through an unsupervised-learning analysis.

Materials and Methods

The analysis was conducted on retrospective data of 1,259 head and neck cancer (HNC) patients treated at the University of Texas MD Anderson Cancer Center between 2005 and 2015. The (structural) clusters of mandibular dose-volume histograms (DVHs) were identified through the K-means clustering method. A soft-margin support vector machine (SVM) was used to determine the cluster borders and partition the dose-volume space. The risk of ORN for each dose-volume region was calculated based on the clinical risk factors and incidence rates.

Results

The K-means clustering method identified six clusters among the DVHs. Based on the first five clusters, the dose-volume space was partitioned almost perfectly by the soft-margin SVM into distinct regions with different risk indices. The sixth cluster overlapped the others entirely; the region of this cluster was determined by its envelops. These regions and the associated risk indices provide a range of constraints for dose optimization under different risk levels.

Conclusion

This study presents an unsupervised-learning analysis of a large-scale data set to evaluate the risk of mandibular ORN among HNC patients. The results provide a visual risk-assessment tool (based on the whole DVH) and a spectrum of dose constraints for radiation planning.

Artificial intelligence to predict outcomes of head and neck radiotherapy

Head and neck radiotherapy induces important toxicity, and its efficacy and tolerance vary widely across patients. Advancements in radiotherapy delivery techniques, along with the increased quality and frequency of image guidance, offer a unique opportunity to individualize radiotherapy based on imaging biomarkers, with the aim of improving radiation efficacy while reducing its toxicity. Various artificial intelligence models integrating clinical data and radiomics have shown encouraging results for toxicity and cancer control outcomes prediction in head and neck cancer radiotherapy. Clinical implementation of these models could lead to individualized risk-based therapeutic decision making, but the reliability of the current studies is limited. Understanding, validating and expanding these models to larger multi-institutional data sets and testing them in the context of clinical trials is needed to ensure safe clinical implementation. This review summarizes the current state of the art of machine learning models for prediction of head and neck cancer radiotherapy outcomes.

Machine learning for normal tissue complication probability prediction: Predictive power with versatility and easy implementation

Background and purpose

A popular Normal tissue Complication (NTCP) model deployed to predict radiotherapy (RT) toxicity is the Lyman-Burman Kutcher (LKB) model of tissue complication. Despite the LKB model's popularity, it can suffer from numerical instability and considers only the generalized mean dose (GMD) to an organ. Machine learning (ML) algorithms can potentially offer superior predictive power of the LKB model, and with fewer drawbacks. Here we examine the numerical characteristics and predictive power of the LKB model and compare these with those of ML.

Materials and methods

Both an LKB model and ML models were used to predict G2 Xerostomia on patients following RT for head and neck cancer, using the dose volume histogram of parotid glands as the input feature. Model speed, convergence characteristics and predictive power was evaluated on an independent training set.

Results

We found that only global optimization algorithms could guarantee a convergent and predictive LKB model. At the same time our results showed that ML models remained unconditionally convergent and predictive, while staying robust to gradient descent optimization. ML models outperform LKB in Brier score and accuracy but compare to LKB in ROC-AUC.

Conclusion

We have demonstrated that ML models can quantify NTCP better than or as well as LKB models, even for a toxicity that the LKB model is particularly well suited to predict. ML models can offer this performance while offering fundamental advantages in model convergence, speed, and flexibility, and so could offer an alternative to the LKB model that could potentially be used in clinical RT planning decisions.

Head and Neck Radiation Therapy Patterns of Practice Variability Identified as a Challenge to Real-World Big Data: Results From the Learning from Analysis of Multicentre Big Data Aggregation (LAMBDA) Consortium

Purpose

Outside of randomized clinical trials, it is difficult to develop clinically relevant evidence-based recommendations for radiation therapy (RT) practice guidelines owing to lack of comprehensive real-world data. To address this knowledge gap, we formed the Learning from Analysis of Multicenter Big Data Aggregation consortium to cooperatively implement RT data standardization, develop software solutions for data analysis, and recommend clinical practice change based on real-world data analyzed. The first phase of this "Big Data" study aimed at characterizing variability in clinical practice patterns of dosimetric data for organs at risk (OARs) that would undermine subsequent use of large-scale, electronically aggregated data to characterize associations with outcomes. Evidence from this study was used as the basis for practical recommendations to improve data quality.

Methods and Materials

Dosimetric details of patients with head and neck cancer treated with radiation therapy between 2014 and 2019 were analyzed. Institutional patterns of practice were characterized, including structure nomenclature, volumes, and frequency of contouring. Dose volume histogram (DVH) distributions were characterized and compared with institutional constraints and literature values.

Results

Plans for 4664 patients treated to a mean plan dose of 64.4 ± 13.2 Gy in 32 ± 4 fractions were aggregated. Before implementation of TG-263 guidelines in each institution, there was variability in OAR nomenclature across institutions and structures. With evidence from this study, we identified a targeted and practical set of recommendations aimed at improving the quality of real-world data.

Conclusions

Quantifying similarities and differences among institutions for OAR structures and DVH metrics is the launching point for next steps to investigate potential relationships between DVH parameters and patient outcomes.

Development of a normal tissue complication probability (NTCP) model using an artificial neural network for radiation-induced necrosis after carbon ion re-irradiation in locally recurrent nasopharyngeal carcinoma

Background

The aim of the present study was to build a normal tissue complication probability (NTCP) model using an artificial neural network (ANN) for radiation-induced necrosis after carbon ion re-irradiation in locally recurrent nasopharyngeal carcinoma (rNPC), and to determine the predictive parameters applied to the model.

Methods

A total of 150 patients with rNPC treated at Shanghai Proton and Heavy Ion Center during 2015-2019 were selected to determine the dominant factors causing mucosal necrosis after carbon therapy. An ANN was built to study both dose-volume histogram (DVH) and clinical factors. Simple oversampling and data normalization were used in the training process. Ten-fold cross validation was conducted to prevent overfitting.

Results

Of the DVH factors, the prediction accuracy ranged from 58.3-65.2%, whereas planning target volume (PTV) receiving dose more than 25 GyE (PTV.V25) yielded the best prediction accuracy. Of the clinical factors, baseline necrosis, sex, and biologically equivalent dose (BED) of initial treatment could increase the accuracy of PTV.V25 by 0.5%, 0.5%, and 1.5%, respectively.

Conclusions

An ANN was built to predict radiation-induced necrosis after re-irradiation in rNPC. The best accuracy and area under receiver-operating characteristic (ROC) curve (AUC) were 66.7% and 0.689. The most predictive dosimetric and clinical parameters were PTV.V25 and BED of initial treatment.

Development and optimisation of a machine-learning prediction model for acute desquamation following breast radiotherapy in the multi-centre REQUITE cohort

Purpose

Some patients with breast cancer treated by surgery and radiation therapy experience clinically significant toxicity, which may adversely affect cosmesis and quality of life. There is a paucity of validated clinical prediction models for radiation toxicity. We used machine learning (ML) algorithms to develop and optimise a clinical prediction model for acute breast desquamation after whole breast external beam radiation therapy in the prospective multicenter REQUITE cohort study.

Methods and Materials

Using demographic and treatment-related features (m = 122) from patients (n = 2058) at 26 centers, we trained 8 ML algorithms with 10-fold cross-validation in a 50:50 random-split data set with class stratification to predict acute breast desquamation. Based on performance in the validation data set, the logistic model tree, random forest, and naïve Bayes models were taken forward to cost-sensitive learning optimisation.

Results

One hundred and ninety-two patients experienced acute desquamation. Resampling and cost-sensitive learning optimisation facilitated an improvement in classification performance. Based on maximising sensitivity (true positives), the “hero” model was the cost-sensitive random forest algorithm with a false-negative: false-positive misclassification penalty of 90:1 containing m = 114 predictive features. Model sensitivity and specificity were 0.77 and 0.66, respectively, with an area under the curve of 0.77 in the validation cohort.

Conclusions

ML algorithms with resampling and cost-sensitive learning generated clinically valid prediction models for acute desquamation using patient demographic and treatment features. Further external validation and inclusion of genomic markers in ML prediction models are worthwhile, to identify patients at increased risk of toxicity who may benefit from supportive intervention or even a change in treatment plan.

Machine Learning for Head and Neck Cancer: A Safe Bet?-A Clinically Oriented Systematic Review for the Radiation Oncologist

BACKGROUND AND PURPOSE

Machine learning (ML) is emerging as a feasible approach to optimize patients' care path in Radiation Oncology. Applications include autosegmentation, treatment planning optimization, and prediction of oncological and toxicity outcomes. The purpose of this clinically oriented systematic review is to illustrate the potential and limitations of the most commonly used ML models in solving everyday clinical issues in head and neck cancer (HNC) radiotherapy (RT).

MATERIALS AND METHODS

Electronic databases were screened up to May 2021. Studies dealing with ML and radiomics were considered eligible. The quality of the included studies was rated by an adapted version of the qualitative checklist originally developed by Luo et al. All statistical analyses were performed using R version 3.6.1.

RESULTS

Forty-eight studies (21 on autosegmentation, four on treatment planning, 12 on oncological outcome prediction, 10 on toxicity prediction, and one on determinants of postoperative RT) were included in the analysis. The most common imaging modality was computed tomography (CT) (40%) followed by magnetic resonance (MR) (10%). Quantitative image features were considered in nine studies (19%). No significant differences were identified in global and methodological scores when works were stratified per their task (i.e., autosegmentation).

DISCUSSION AND CONCLUSION

The range of possible applications of ML in the field of HN Radiation Oncology is wide, albeit this area of research is relatively young. Overall, if not safe yet, ML is most probably a bet worth making.

A systematically compiled set of quantitative metrics to describe spatial characteristics of radiotherapy dose distributions and aid in treatment planning

PURPOSE

Many quantitative metrics have been proposed in literature for characterization of spatial dose properties. The aim of this study is to work towards much-needed consensus in the radiotherapy community on which of these metrics to use. We do this by comparing characteristics of the metrics and providing a systematically selected set of metrics to comprehensively quantify properties of the spatial dose distribution.

METHODS

We searched the literature for metrics to quantitatively evaluate dose conformity, homogeneity, gradient (overall and directional), and distribution and location of over- and under-dosed sub-volumes. For each spatial dose property, we compared the responses of its corresponding metrics to simulated dose variations in a virtual water phantom. Selection criteria were a metric's ability to describe simulated scenarios robustly and to be visualized in an intuitive way.

RESULTS

We saw substantial differences in the responses of metrics to the simulated dose variations. Some conformity and homogeneity metrics were unable to quantify certain types of changes (e.g. target under-coverage). Others showed a large dependency on the shape and volume of targets and isodoses. Metric values differed between calculations in a static plan and in simulated full treatment courses including setup errors, especially for metrics quantifying distribution and location of hot and cold spots. We provide an Eclipse plugin script to calculate and visualize selected metrics.

CONCLUSION

The selected set of metrics provides complementary and comprehensive quantitative information about the spatial dose distribution. This work serves as a step towards broader consensus on the use of spatial dose metrics.

Spatial descriptions of radiotherapy dose: normal tissue complication models and statistical associations

For decades, dose-volume information for segmented anatomy has provided the essential data for correlating radiotherapy dosimetry with treatment-induced complications. Dose-volume information has formed the basis for modelling those associations via normal tissue complication probability (NTCP) models and for driving treatment planning. Limitations to this approach have been identified. Many studies have emerged demonstrating that the incorporation of information describing the spatial nature of the dose distribution, and potentially its correlation with anatomy, can provide more robust associations with toxicity and seed more general NTCP models. Such approaches are culminating in the application of computationally intensive processes such as machine learning and the application of neural networks. The opportunities these approaches have for individualising treatment, predicting toxicity and expanding the solution space for radiation therapy are substantial and have clearly widespread and disruptive potential. Impediments to reaching that potential include issues associated with data collection, model generalisation and validation. This review examines the role of spatial models of complication and summarises relevant published studies. Sources of data for these studies, appropriate statistical methodology frameworks for processing spatial dose information and extracting relevant features are described. Spatial complication modelling is consolidated as a pathway to guiding future developments towards effective, complication-free radiotherapy treatment.

Comparison of machine learning methods for prediction of osteoradionecrosis incidence in patients with head and neck cancer

OBJECTIVES

Mandible osteoradionecrosis (ORN) is one of the most severe toxicities in patients with head and neck cancer (HNC) undergoing radiotherapy (RT). The existing literature focuses on the correlation of mandible ORN and clinical and dosimetric factors. This study proposes the use of machine learning (ML) methods as prediction models for mandible ORN incidence.

METHODS

A total of 96 patients (ORN incidence ratio of 1:1) treated between 2011 and 2015 were selected from the local HNC toxicity database. Demographic, clinical and dosimetric data (based on the mandible dose-volume histogram) were considered as model variables. Prediction accuracy (measured using a stratified fivefold nested cross-validation), sensitivity, specificity, precision and negative predictive value were used to evaluate the prediction performance of a multivariate logistic regression (LR) model, a support vector machine (SVM) model, a random forest (RF) model, an adaptive boosting (AdaBoost) model and an artificial neural network (ANN) model. The different models were compared based on their prediction accuracy and using the McNemar's hypothesis test.

RESULTS

The ANN model (77% accuracy), closely followed by the SVM (76%), AdaBoost (75%) and LR (75%) models, showed the highest overall prediction accuracy. The RF model (71%) showed the lowest prediction accuracy. However, based on the McNemar's test applied to all model pair combinations, no statistically significant difference between the models was found.

CONCLUSION

Based on our results, we encourage the use of ML-based prediction models for ORN incidence as has already been done for other HNC toxicity end points.

ADVANCES IN KNOWLEDGE

This research opens a new path towards personalised RT for HNC using ML to predict mandible ORN incidence.

Artificial intelligence and machine learning for medical imaging: A technology review

Artificial intelligence (AI) has recently become a very popular buzzword, as a consequence of disruptive technical advances and impressive experimental results, notably in the field of image analysis and processing. In medicine, specialties where images are central, like radiology, pathology or oncology, have seized the opportunity and considerable efforts in research and development have been deployed to transfer the potential of AI to clinical applications. With AI becoming a more mainstream tool for typical medical imaging analysis tasks, such as diagnosis, segmentation, or classification, the key for a safe and efficient use of clinical AI applications relies, in part, on informed practitioners. The aim of this review is to present the basic technological pillars of AI, together with the state-of-the-art machine learning methods and their application to medical imaging. In addition, we discuss the new trends and future research directions. This will help the reader to understand how AI methods are now becoming an ubiquitous tool in any medical image analysis workflow and pave the way for the clinical implementation of AI-based solutions.

Development and application of an elastic net logistic regression model to investigate the impact of cardiac substructure dose on radiation-induced pericardial effusion in patients with NSCLC

BACKGROUND

Typically, cardiac substructures are neither delineated nor analyzed during radiation treatment planning. Therefore, we developed a novel machine learning model to evaluate the impact of cardiac substructure dose for predicting radiation-induced pericardial effusion (PCE).

MATERIALS AND METHODS

One-hundred and forty-one stage III NSCLC patients, who received radiation therapy in a prospective clinical trial, were included in this analysis. The impact of dose-volume histogram (DVH) metrics (mean and max dose, V5Gy[%]-V70Gy[%]) for the whole heart, left and right atrium, and left and right ventricle, on pericardial effusion toxicity (≥grade 2, CTCAE v4.0 grading) were examined. Elastic net logistic regression, using repeat cross-validation (n = 100 iterations, 75%/25% training/test set data split), was conducted with cardiac-based DVH metrics as covariates. The following model types were constructed and analyzed: (i) standard model type, which only included whole-heart DVH metrics; and (ii) a model type trained with both whole-heart and substructure DVH metrics. Model performance was analyzed on the test set using area under the curve (AUC), accuracy, calibration slope and calibration intercept. A final fitted model, based on the optimal model type, was developed from the entire study population for future comparisons.

RESULTS

Grade 2 PCE incidence was 49.6% (n = 70). Models using whole heart and substructure dose had the highest performance (median values: AUC = 0.820; calibration slope/intercept = 1.356/-0.235; accuracy = 0.743) and outperformed the standard whole-heart only model type (median values: AUC = 0.799; calibration slope/intercept = 2.456/-0.729; accuracy = 0.713). The final fitted elastic net model showed high performance in predicting PCE (median values: AUC = 0.879; calibration slope/intercept = 1.352/-0.174; accuracy = 0.801).

CONCLUSIONS

We developed and evaluated elastic net regression toxicity models of radiation-induced PCE. We found the model type that included cardiac substructure dose had superior predictive performance. A final toxicity model that included cardiac substructure dose metrics was developed and reported for comparison with external datasets.

Artificial intelligence in radiation oncology

Artificial intelligence (AI) has the potential to fundamentally alter the way medicine is practised. AI platforms excel in recognizing complex patterns in medical data and provide a quantitative, rather than purely qualitative, assessment of clinical conditions. Accordingly, AI could have particularly transformative applications in radiation oncology given the multifaceted and highly technical nature of this field of medicine with a heavy reliance on digital data processing and computer software. Indeed, AI has the potential to improve the accuracy, precision, efficiency and overall quality of radiation therapy for patients with cancer. In this Perspective, we first provide a general description of AI methods, followed by a high-level overview of the radiation therapy workflow with discussion of the implications that AI is likely to have on each step of this process. Finally, we describe the challenges associated with the clinical development and implementation of AI platforms in radiation oncology and provide our perspective on how these platforms might change the roles of radiotherapy medical professionals.

Machine Learning-Based Models for Prediction of Toxicity Outcomes in Radiotherapy

In order to limit radiotherapy (RT)-related side effects, effective toxicity prediction and assessment schemes are essential. In recent years, the growing interest toward artificial intelligence and machine learning (ML) within the science community has led to the implementation of innovative tools in RT. Several researchers have demonstrated the high performance of ML-based models in predicting toxicity, but the application of these approaches in clinics is still lagging, partly due to their low interpretability. Therefore, an overview of contemporary research is needed in order to familiarize practitioners with common methods and strategies. Here, we present a review of ML-based models for predicting and classifying RT-induced complications from both a methodological and a clinical standpoint, focusing on the type of features considered, the ML methods used, and the main results achieved. Our work overviews published research in multiple cancer sites, including brain, breast, esophagus, gynecological, head and neck, liver, lung, and prostate cancers. The aim is to define the current state of the art and main achievements within the field for both researchers and clinicians.

Dose cluster model parameterization of the parotid gland in irradiation of head and neck cancer

To explore the parotid normal tissue complication probability (NTCP) modeling with percolation-based dose clusters for head-and-neck patients receiving concomitant chemotherapy and radiation therapy. Cluster models incorporating the spatial dose distribution in the parotid gland were developed to evaluate the radiation induced complication. Cluster metrics including the mean cluster size (NMCS) and the largest cluster size both normalized by the gland volume (NSLC) were evaluated and scrutinized against the benchmark NTCP. Two fitting strategies to the Lyman-Kutcher-Burman (LKB) model using the maximum likelihood method were devised: the volume parameter n fixed at 1.0 (mean dose model) and unrestricted (full LKB model). The fitted parameters TD50 and m were assessed with the LKB NTCP models with the available xerostomia data. NSLC was a better metric than NMCS with reference to the LKB model and strong correlation (r ~ 0.95) was observed between NTCP and NSLC. The mean dose model returned the parameter TD50 (39.9 Gy) and m (0.4) from the NSLC of threshold dose at around 40 Gy. Drastically different TD50 and m values were obtained from the fittings via the full LKB model, where the threshold dose would be near 27 Gy. Bootstrapping analyses further confirmed the fitting outcomes. Strong correlation with the traditional NTCP models revealed that the cluster model could achieve what NTCP models attain and may offer additional information. Parameterization of the model indicated that the model could have different predictions from current clinical recommendations. Further investigation using toxicity data is under way to validate the cluster model.

Predicting Emergency Visits and Hospital Admissions During Radiation and Chemoradiation: An Internally Validated Pretreatment Machine Learning Algorithm

PURPOSE

Patients undergoing radiotherapy (RT) or chemoradiotherapy (CRT) may require emergency department evaluation or hospitalization. Early identification may direct preventative supportive care, improving outcomes and reducing health care costs. We developed and evaluated a machine learning (ML) approach to predict these events.

METHODS

A total of 8,134 outpatient courses of RT and CRT from a single institution from 2013 to 2016 were identified. Extensive pretreatment data were programmatically extracted and processed from the electronic health record (EHR). Training and internal validation cohorts were randomly generated (3:1 ratio). Gradient tree boosting (GTB), random forest, support vector machine, and least absolute shrinkage and selection operator logistic regression approaches were trained and internally validated based on area under receiver operating characteristic (AUROC) curve. The most predictive ML approach was also evaluated using only disease- and treatment-related factors to assess predictive gain of extensive EHR data.

RESULTS

All methods had high predictive accuracy, particularly GTB (validation AUROC, 0.798). Extensive EHR data beyond disease and treatment information improved accuracy (delta AUROC, 0.056). A Youden-based cutoff corresponded to validation sensitivity of 81.0% (175 of 216 courses with events) and specificity of 67.3% (1,218 of 1811 courses without events). Interpretability is an important advantage of GTB. Variable importance identified top predictive factors, including treatment (planned RT and systemic therapy), pretreatment encounters (emergency department visits and admissions in the year before treatment), vital signs (weight loss and pain score in the year before treatment), and laboratory values (albumin level at weeks before treatment).

CONCLUSION

ML predicts emergency visits and hospitalization during cancer therapy. Incorporating predictions into clinical care algorithms may help direct personalized supportive care, improve quality of care, and reduce costs. A prospective trial investigating ML-assisted direction of increased clinical assessments during RT is planned.

Incorporation of Dosimetric Gradients and Parotid Gland Migration Into Xerostomia Prediction

Purpose: Due to the sharp gradients of intensity-modulated radiotherapy (IMRT) dose distributions, treatment uncertainties may induce substantial deviations from the planned dose during irradiation. Here, we investigate if the planned mean dose to parotid glands in combination with the dose gradient and information about anatomical changes during the treatment improves xerostomia prediction in head and neck cancer patients.

Materials and methods: Eighty eight patients were retrospectively analyzed. Three features of the contralateral parotid gland were studied in terms of their association with the outcome, i.e., grade ≥ 2 (G2) xerostomia between 6 months and 2 years after radiotherapy (RT): planned mean dose (MD), average lateral dose gradient (GRADX), and parotid gland migration toward medial (PGM). PGM was estimated using daily megavoltage computed tomography (MVCT) images. Three logistic regression models where analyzed: based on (1) MD only, (2) MD and GRADX, and (3) MD, GRADX, and PGM. Additionally, the cohort was stratified based on the median value of GRADX, and a univariate analysis was performed to study the association of the MD with the outcome for patients in low- and high-GRADX domains.

Results: The planned MD failed to recognize G2 xerostomia patients (AUC = 0.57). By adding the information of GRADX (second model), the model performance increased to AUC = 0.72. The addition of PGM (third model) led to further improvement in the recognition of the outcome (AUC = 0.79). Remarkably, xerostomia patients in the low-GRADX domain were successfully identified (AUC = 0.88) by the MD alone.

Conclusions: Our results indicate that GRADX and PGM, which together serve as a proxy of dosimetric changes, provide valuable information for xerostomia prediction.

PACE: A Probabilistic Atlas for Normal Tissue Complication Estimation in Radiation Oncology

In radiation oncology, the need for a modern Normal Tissue Complication Probability (NTCP) philosophy to include voxel-based evidence on organ radio-sensitivity (RS) has been acknowledged. Here a new formalism (Probabilistic Atlas for Complication Estimation, PACE) to predict radiation-induced morbidity (RIM) is presented. The adopted strategy basically consists in keeping the structure of a classical, phenomenological NTCP model, such as the Lyman-Kutcher-Burman (LKB), and replacing the dose distribution with a collection of RIM odds, including also significant non-dosimetric covariates, as input of the model framework. The theory was first demonstrated in silico on synthetic dose maps, classified according to synthetic outcomes. PACE was then applied to a clinical dataset of thoracic cancer patients classified for lung fibrosis. LKB models were trained for comparison. Overall, the obtained learning curves showed that the PACE model outperformed the LKB and predicted synthetic outcomes with an accuracy >0.8. On the real patients, PACE performance, evaluated by both discrimination and calibration, was significantly higher than LKB. This trend was confirmed by cross-validation. Furthermore, the capability to infer the spatial pattern of underlying RS map for the analyzed RIM was successfully demonstrated, thus paving the way to new perspectives of NTCP models as learning tools.

Image-based Data Mining to Probe Dosimetric Correlates of Radiation-induced Trismus

PURPOSE

To identify imaged regions in which dose is associated with radiation-induced trismus after head and neck cancer radiation therapy (HNRT) using a novel image-based data mining (IBDM) framework.

METHODS AND MATERIALS

A cohort of 86 HNRT patients were analyzed for region identification. Trismus was characterized as a continuous variable by the maximum incisor-to-incisor opening distance (MID) at 6 months after radiation therapy. Patient anatomies and dose distributions were spatially normalized to a common frame of reference using deformable image registration. IBDM was used to identify clusters of voxels associated with MID (P ≤ .05 based on permutation testing). The result was externally tested on a cohort of 35 patients with head and neck cancer. Internally, we also performed a dose-volume histogram-based analysis by comparing the magnitude of the correlation between MID and the mean dose for the IBDM-identified cluster in comparison with 5 delineated masticatory structures.

RESULTS

A single cluster was identified with the IBDM approach (P < .01), partially overlapping with the ipsilateral masseter. The dose-volume histogram-based analysis confirmed that the IBDM cluster had the strongest association with MID, followed by the ipsilateral masseter and the ipsilateral medial pterygoid (Spearman's rank correlation coefficients: R_s = -0.36, -0.35, -0.32; P = .001, .001, .002, respectively). External validation confirmed an association between mean dose to the IBDM cluster and MID (R_s = -0.45; P = .007).

CONCLUSIONS

IBDM bypasses the common assumption that dose patterns within structures are unimportant. Our novel IBDM approach for continuous outcome variables successfully identified a cluster of voxels that are highly associated with trismus, overlapping partially with the ipsilateral masseter. Tests on an external validation cohort showed an even stronger correlation with trismus. These results support use of the region in HNRT treatment planning to potentially reduce trismus.