Functional Data Analysis Applied to Modeling of Severe Acute Mucositis and Dysphagia Resulting From Head and Neck Radiation Therapy

Performance Comparison of 10 State-of-the-Art Machine Learning Algorithms for Outcome Prediction Modeling of Radiation-Induced Toxicity

Purpose

To evaluate the efficacy of prominent machine learning algorithms in predicting normal tissue complication probability using clinical data obtained from 2 distinct disease sites and to create a software tool that facilitates the automatic determination of the optimal algorithm to model any given labeled data set.

Methods and Materials

We obtained 3 sets of radiation toxicity data (478 patients) from our clinic: gastrointestinal toxicity, radiation pneumonitis, and radiation esophagitis. These data comprised clinicopathological and dosimetric information for patients diagnosed with non-small cell lung cancer and anal squamous cell carcinoma. Each data set was modeled using 11 commonly employed machine learning algorithms (elastic net, least absolute shrinkage and selection operator [LASSO], random forest, random forest regression, support vector machine, extreme gradient boosting, light gradient boosting machine, k-nearest neighbors, neural network, Bayesian-LASSO, and Bayesian neural network) by randomly dividing the data set into a training and test set. The training set was used to create and tune the model, and the test set served to assess it by calculating performance metrics. This process was repeated 100 times by each algorithm for each data set. Figures were generated to visually compare the performance of the algorithms. A graphical user interface was developed to automate this whole process.

Results

LASSO achieved the highest area under the precision-recall curve (0.807 ± 0.067) for radiation esophagitis, random forest for gastrointestinal toxicity (0.726 ± 0.096), and the neural network for radiation pneumonitis (0.878 ± 0.060). The area under the curve was 0.754 ± 0.069, 0.889 ± 0.043, and 0.905 ± 0.045, respectively. The graphical user interface was used to compare all algorithms for each data set automatically. When averaging the area under the precision-recall curve across all toxicities, Bayesian-LASSO was the best model.

Conclusions

Our results show that there is no best algorithm for all data sets. Therefore, it is important to compare multiple algorithms when training an outcome prediction model on a new data set. The graphical user interface created for this study automatically compares the performance of these 11 algorithms for any data set.

A marginal structural model for normal tissue complication probability

The goal of radiation therapy for cancer is to deliver prescribed radiation dose to the tumor while minimizing dose to the surrounding healthy tissues. To evaluate treatment plans, the dose distribution to healthy organs is commonly summarized as dose-volume histograms (DVHs). Normal tissue complication probability (NTCP) modeling has centered around making patient-level risk predictions with features extracted from the DVHs, but few have considered adapting a causal framework to evaluate the safety of alternative treatment plans. We propose causal estimands for NTCP based on deterministic and stochastic interventions, as well as propose estimators based on marginal structural models that impose bivariable monotonicity between dose, volume, and toxicity risk. The properties of these estimators are studied through simulations, and their use is illustrated in the context of radiotherapy treatment of anal canal cancer patients.

A Bayesian joint model for mediation analysis with matrix-valued mediators

Unscheduled treatment interruptions may lead to reduced quality of care in radiation therapy (RT). Identifying the RT prescription dose effects on the outcome of treatment interruptions, mediated through doses distributed into different organs at risk (OARs), can inform future treatment planning. The radiation exposure to OARs can be summarized by a matrix of dose-volume histograms (DVH) for each patient. Although various methods for high-dimensional mediation analysis have been proposed recently, few studies investigated how matrix-valued data can be treated as mediators. In this paper, we propose a novel Bayesian joint mediation model for high-dimensional matrix-valued mediators. In this joint model, latent features are extracted from the matrix-valued data through an adaptation of probabilistic multilinear principal components analysis (MPCA), retaining the inherent matrix structure. We derive and implement a Gibbs sampling algorithm to jointly estimate all model parameters, and introduce a Varimax rotation method to identify active indicators of mediation among the matrix-valued data. Our simulation study finds that the proposed joint model has higher efficiency in estimating causal decomposition effects compared to an alternative two-step method, and demonstrates that the mediation effects can be identified and visualized in the matrix form. We apply the method to study the effect of prescription dose on treatment interruptions in anal canal cancer patients.

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 that incorporates the whole radiation dose distribution on the mandible.

METHODS AND MATERIALS

The analysis was conducted on retrospective data of 1259 patients with head and neck cancer treated at The University of Texas MD Anderson Cancer Center between 2005 and 2015. During a minimum 12-month posttherapy follow-up period, 173 patients in this cohort (13.7%) developed ORN (grades I to IV). The (structural) clusters of mandibular dose-volume histograms (DVHs) for these patients were identified using the K-means clustering method. A soft-margin support vector machine 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 incidence rates and other clinical risk factors.

RESULTS

The K-means clustering method identified 6 clusters among the DVHs. Based on the first 5 clusters, the dose-volume space was partitioned by the soft-margin support vector machine into distinct regions with different risk indices. The sixth cluster entirely overlapped with the others; the region of this cluster was determined by its envelopes. For each region, the ORN incidence rate per preradiation dental extraction status (a statistically significant, nondose related risk factor for ORN) was reported as the corresponding risk index.

CONCLUSIONS

This study presents an unsupervised-learning analysis of a large-scale data set to evaluate the risk of mandibular ORN among patients with head and neck cancer. The results provide a visual risk-assessment tool for ORN (based on the whole DVH and preradiation dental extraction status) as well as a range of constraints for dose optimization under different risk levels.

Declarations of Independence: How Embedded Multicollinearity Errors Affect Dosimetric and Other Complex Analyses in Radiation Oncology

The statistical technique of multiple regression, commonly referred to as "multivariable regression," is often used in clinical research to quantify the relationships between multiple predictor variables and a single outcome variable of interest. The foundational theory underpinning multivariable regression assumes that all predictor variables are independent of one another. In other words, the effect of each independent variable is measured by its contribution to the regression equation while all other variables remain unchanged. In the presence of correlations between two or more variables, however, it is impossible to change one variable without a consequent change in the variable(s) it is linked to. This condition, known as "multicollinearity," can introduce errors into multivariable regression models by affecting estimates of the regression coefficients that quantify the relationship between individual predictor variables and the outcome variable. Errors that arise due to violations of the multicollinearity assumption are of special interest to radiation oncology researchers. Because of high levels of correlation among variables derived from points along individual organ dose-volume histogram (DVH) curves, as well as strong intercorrelations among dose-volume parameters in neighboring organs, dosimetric analyses are particularly subject to multicollinearity errors. For example, dose-volume parameters for the heart are strongly correlated not only with other points along the heart DVH curve but are likely also correlated with dose-volume parameters in neighboring organs such as the lung. In this paper, we describe the problem of multicollinearity in accessible terms and discuss examples of violations of the multicollinearity assumption within the radiation oncology literature. Finally, we provide recommendations regarding best practices for identifying and managing multicollinearity in complex data sets.

Proton therapy and oral mucositis in oral & oropharyngeal cancers: outcomes, dosimetric and NTCP benefit

INTRODUCTION

Radiation-induced oral mucositis (RIOM), is a common, debilitating, acute side effect of radiotherapy for oral cavity (OC) and oropharyngeal (OPx) cancers; technical innovations for reducing it are seldom discussed. Intensity-modulated-proton-therapy (IMPT) has been reported extensively for treating OPx cancers, and less frequently for OC cancers. We aim to quantify the reduction in the likelihood of RIOM in treating these 2 subsites with IMPT compared to Helical Tomotherapy.

MATERIAL AND METHODS

We report acute toxicities and early outcomes of 22 consecutive patients with OC and OPx cancers treated with IMPT, and compare the dosimetry and normal tissue complication probability (NTCP) of ≥ grade 3 mucositis for IMPT and HT.

RESULTS

Twenty two patients, 77% males, 41% elderly and 73% OC subsite, were reviewed. With comparable target coverage, IMPT significantly reduced the mean dose and D32, D39, D45, and D50, for both the oral mucosa (OM) and spared oral mucosa (sOM). With IMPT, there was a 7% absolute and 16.5% relative reduction in NTCP for grade 3 mucositis for OM, compared to HT. IMPT further reduced NTCP for sOM, and the benefit was maintained in OC, OPx subsites and elderly subgroup. Acute toxicities, grade III dermatitis and mucositis, were noted in 50% and 45.5% patients, respectively, while 22.7% patients had grade 3 dysphagia. Compared with published data, the hospital admission rate, median weight loss, feeding tube insertion, unplanned treatment gaps were lower with IMPT. At a median follow-up of 15 months, 81.8% were alive; 72.7%, alive without disease and 9%, alive with disease.

CONCLUSION

The dosimetric benefit of IMPT translates into NTCP reduction for grade 3 mucositis compared to Helical Tomotherapy for OPx and OC cancers and encourages the use of IMPT in their management.

Functional principal component analysis for identifying the child growth pattern using longitudinal birth cohort data

Background

Longitudinal studies are important to understand patterns of growth in children and limited in India. It is important to identify an approach for characterising growth trajectories to distinguish between children who have healthy growth and those growth is poor. Many statistical approaches are available to assess the longitudinal growth data and which are difficult to recognize the pattern. In this research study, we employed functional principal component analysis (FPCA) as a statistical method to find the pattern of growth data. The purpose of this study is to describe the longitudinal child growth trajectory pattern under 3 years of age using functional principal component method.

Methods

Children born between March 2002 and August 2003 (n = 290) were followed until their third birthday in three neighbouring slums in Vellore, South India. Field workers visited homes to collect details of morbidity twice a week. Height and weight were measured monthly from 1 month of age in a study-run clinic. Longitudinal child growth trajectory pattern were extracted using Functional Principal Component analysis using B-spline basis functions with smoothing parameters. Functional linear model was used to assess the factors association with the growth functions.

Results

We have obtained four FPCs explained by 86.5, 3.9, 3.1 and 2.2% of the variation respectively for the height functions. For height, 38% of the children’s had poor growth trajectories. Similarly, three FPCs explained 76.2, 8.8, and 4.7% respectively for the weight functions and 44% of the children’s had poor growth in their weight trajectories. Results show that gender, socio-economic status, parent’s education, breast feeding, and gravida are associated and, influence the growth pattern in children.

Conclusions

The FPC approach deals with subjects’ dynamics of growth and not with specific values at given times. FPC could be a better alternate approach for both dimension reduction and pattern detection. FPC may be used to offer greater insight for classification.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12874-022-01566-0.

Classification of tolerable/intolerable mucosal toxicity of head-and-neck radiotherapy schedules with a biomathematical model of cell dynamics

Purpose

The purpose of this study is to present a biomathematical model based on the dynamics of cell populations to predict the tolerability/intolerability of mucosal toxicity in head‐and‐neck radiotherapy.

Methods and Materials

Our model is based on the dynamics of proliferative and functional cell populations in irradiated mucosa, and incorporates the three As: Accelerated proliferation, loss of Asymmetric proliferation, and Abortive divisions. The model consists of a set of delay differential equations, and tolerability is based on the depletion of functional cells during treatment. We calculate the sensitivity (sen) and specificity (spe) of the model in a dataset of 108 radiotherapy schedules, and compare the results with those obtained with three phenomenological classification models, two based on a biologically effective dose (BED) function describing the tolerability boundary (Fowler and Fenwick) and one based on an equivalent dose in 2 Gy fractions (EQD₂) boundary (Strigari). We also perform a machine learning‐like cross‐validation of all the models, splitting the database in two, one for training and one for validation.

Results

When fitting our model to the whole dataset, we obtain predictive values (sen + spe) up to 1.824. The predictive value of our model is very similar to that of the phenomenological models of Fowler (1.785), Fenwick (1.806), and Strigari (1.774). When performing a k = 2 cross‐validation, the specificity and sensitivity in the validation dataset decrease for all models, from ˜1.82 to ˜1.55–1.63. For Fowler, the worsening is higher, down to 1.49.

Conclusions

Our model has proved useful to predict the tolerability/intolerability of a dataset of 108 schedules. As the model is more mechanistic than other available models, it could prove helpful when designing unconventional dose fractionations, schedules not covered by datasets to which phenomenological models of toxicity have been fitted.

Early MRI Blood Volume Changes in Constrictor Muscles Correlate With Postradiation Dysphagia

PURPOSE

Predicting individual patient sensitivity to radiation therapy (RT) for tumor control or normal tissue toxicity is necessary to individualize treatment planning. In head and neck cancer, radiation doses are limited by many nearby critical structures, including structures involved in swallowing. Previous efforts showed that imaging parameters correlate with RT dose; here, we investigate the role of dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) blood volume (BV) changes in predicting dysphagia.

METHODS AND MATERIALS

This study included 32 patients with locally advanced oropharyngeal squamous cell carcinoma treated with definitive chemoradiation on an institutional protocol incorporating baseline and early midtreatment DCE-MRI. BV maps of the pharyngeal constrictor muscles (PCM) were created, and BV increases midtreatment were correlated with the following parameters at 3 and 12 months post-RT: RT dose, Dynamic Imaging Grade of Swallowing Toxicity swallow score, aspiration frequency, European Organisation for Research and Treatment of Cancer HN35 patient-reported outcomes, physician-reported dysphagia, and feeding tube (FT) dependence.

RESULTS

The mean BV to the PCMs increased from baseline to fraction 10, which was significant for the superior PCM (P = .006) and middle PCM (P < .001), with a trend in the inferior PCM where lower mean doses were seen (P = .077). The factors associated with FT dependence at 3 months included BV increases in the total PCM (correlation, 0.48; P = .006) and middle PCM (correlation, 0.50; P = .004). A post-RT increase in aspiration was associated with a BV increase in the superior PCM (correlation, 0.44; P = .013),and the increase in the total PCMs was marginally significant (correlation, 0.34; P = .06). The best-performing models of FT dependence (area under the receiver operating curve [AUC] = 0.84) and aspiration increases (AUC = 0.78) included BV increases as well as a mean RT dose to middle PCM.

CONCLUSIONS

Our results suggest that midtreatment BV increases derived from DCE-MRI are an early predictor of dysphagia. Further investigation of these promising imaging markers to assess individual patient sensitivity to treatment and the patient's subsequent risk of toxicities is warranted to improve personalization of RT planning.

A mathematical model of dynamics of cell populations in squamous epithelium after irradiation

PURPOSE

To develop multi-compartment mechanistic models of dynamics of stem and functional cell populations in epithelium after irradiation. Methods and materials: We present two models, with three (3C) and four (4C) compartments respectively. We use delay differential equations, and include accelerated proliferation, loss of division asymmetry, progressive death of abortive stem cells, and turnover of functional cells. The models are used to fit experimental data on the variations of the number of cells in mice mucosa after irradiation with 13 Gy and 20 Gy. Akaike information criteria (AIC) was used to evaluate the performance of each model.

RESULTS

Both 3C and 4C models provide good fits to experimental data for 13 Gy. Fits for 20 Gy are slightly poorer and may be affected by larger uncertainties and fluctuations of experimental data. Best fits are obtained by imposing constraints on the fitting parameters, so to have values that are within experimental ranges. There is some degeneration in the fits, as different sets of parameters provide similarly good fits.

CONCLUSIONS

The models provide good fits to experimental data. Mechanistic approaches like this can facilitate the development of mucositis response models to nonstandard schedules/treatment combinations not covered by datasets to which phenomenological models have been fitted. Studying the dynamics of cell populations in multifraction treatments, and finding links with induced toxicity, is the next step of this work.

Histogram-based models on non-thin section chest CT predict invasiveness of primary lung adenocarcinoma subsolid nodules

109 pathologically proven subsolid nodules (SSN) were segmented by 2 readers on non-thin section chest CT with a lung nodule analysis software followed by extraction of CT attenuation histogram and geometric features. Functional data analysis of histograms provided data driven features (FPC1,2,3) used in further model building. Nodules were classified as pre-invasive (P1, atypical adenomatous hyperplasia and adenocarcinoma in situ), minimally invasive (P2) and invasive adenocarcinomas (P3). P1 and P2 were grouped together (T1) versus P3 (T2). Various combinations of features were compared in predictive models for binary nodule classification (T1/T2), using multiple logistic regression and non-linear classifiers. Area under ROC curve (AUC) was used as diagnostic performance criteria. Inter-reader variability was assessed using Cohen’s Kappa and intra-class coefficient (ICC). Three models predicting invasiveness of SSN were selected based on AUC. First model included 87.5 percentile of CT lesion attenuation (Q.875), interquartile range (IQR), volume and maximum/minimum diameter ratio (AUC:0.89, 95%CI:[0.75 1]). Second model included FPC1, volume and diameter ratio (AUC:0.91, 95%CI:[0.77 1]). Third model included FPC1, FPC2 and volume (AUC:0.89, 95%CI:[0.73 1]). Inter-reader variability was excellent (Kappa:0.95, ICC:0.98). Parsimonious models using histogram and geometric features differentiated invasive from minimally invasive/pre-invasive SSN with good predictive performance in non-thin section CT.

Revisiting the dose constraints for head and neck OARs in the current era of IMRT

Head and neck cancer poses a particular challenge in radiation therapy, whilst being an effective treatment modality it requires very high doses of radiation to provide effective therapy. This is further complicated by the fact that the head and neck region contains a large number of radiosensitive tissues, often resulting in patients experiencing debilitating normal tissue complications. In the era of intensity-modulated radiation therapy (IMRT) treatments can be delivered using non-uniform dose distributions selectively aimed at reducing the dose to critical organs-at-risk while still adequately covering the tumor target. Dose-volume constraints for the different risk organs play a vital role in one's ability to devise the best IMRT treatment plan for a head and neck cancer patient. To this end, it is pivotal to have access to the latest and most relevant dose constraints available and as such the goal of this review is to provide a summary of suggested dose-volume constraints for head and neck cancer RT that have been published after the QUANTEC reports were made available in early 2010.

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

Severe acute dysphagia commonly results from head and neck radiotherapy (RT). A model enabling prediction of severity of acute dysphagia for individual patients could guide clinical decision-making. Statistical associations between RT dose distributions and dysphagia could inform RT planning protocols aiming to reduce the incidence of severe dysphagia. We aimed to establish such a model and associations incorporating spatial dose metrics. Models of severe acute dysphagia were developed using pharyngeal mucosa (PM) RT dose (dose-volume and spatial dose metrics) and clinical data. Penalized logistic regression (PLR), support vector classification and random forest classification (RFC) models were generated and internally (173 patients) and externally (90 patients) validated. These were compared using area under the receiver operating characteristic curve (AUC) to assess performance. Associations between treatment features and dysphagia were explored using RFC models. The PLR model using dose-volume metrics (PLRstandard) performed as well as the more complex models and had very good discrimination (AUC = 0.82) on external validation. The features with the highest RFC importance values were the volume, length and circumference of PM receiving 1 Gy/fraction and higher. The volumes of PM receiving 1 Gy/fraction or higher should be minimized to reduce the incidence of severe acute dysphagia.

Systematic Review of Normal Tissue Complication Models Relevant to Standard Fractionation Radiation Therapy of the Head and Neck Region Published After the QUANTEC Reports

There has recently been an increasing interest in model-based evaluation and comparison of different treatment options in radiation oncology studies. This is partly driven by the considerable technical advancements in radiation therapy of the last decade, leaving radiation oncologists with a multitude of options to consider. In lieu of randomized trials comparing all of these different treatment options for varying indications, which is unfeasible, treatment evaluations based on normal tissue complication probability (NTCP) models offer a practical alternative. The Quantitative Analyses of Normal Tissue Effects in the Clinic (QUANTEC) effort, culminating in a number of reports published in 2010, provided a basis for many of the since-implemented dose-response models and dose-volume constraints and was a key component for model-based treatment evaluations. Given that 7 years have passed since the QUANTEC publications and that patient-reported outcomes have emerged as an important consideration in recent years, an updated summary of the published radiation dose-response literature, which includes a focus on patient-reported quality of life outcomes, is warranted. Here we provide a systematic review of quantitative dose-response models published after January 1, 2010 for endpoints relevant to radiation therapy for head and neck cancer, because these patients are typically at risk for a variety of treatment-induced normal tissue complications.

Parotid gland mean dose as a xerostomia predictor in low-dose domains

PURPOSE

Xerostomia is a common side effect of radiotherapy resulting from excessive irradiation of salivary glands. Typically, xerostomia is modeled by the mean dose-response characteristic of parotid glands and prevented by mean dose constraints to either contralateral or both parotid glands. The aim of this study was to investigate whether normal tissue complication probability (NTCP) models based on the mean radiation dose to parotid glands are suitable for the prediction of xerostomia in a highly conformal low-dose regime of modern intensity-modulated radiotherapy (IMRT) techniques.

MATERIAL AND METHODS

We present a retrospective analysis of 153 head and neck cancer patients treated with radiotherapy. The Lyman Kutcher Burman (LKB) model was used to evaluate predictive power of the parotid gland mean dose with respect to xerostomia at 6 and 12 months after the treatment. The predictive performance of the model was evaluated by receiver operating characteristic (ROC) curves and precision-recall (PR) curves.

RESULTS

Average mean doses to ipsilateral and contralateral parotid glands were 25.4 Gy and 18.7 Gy, respectively. QUANTEC constraints were met in 74% of patients. Mild to severe (G1+) xerostomia prevalence at both 6 and 12 months was 67%. Moderate to severe (G2+) xerostomia prevalence at 6 and 12 months was 20% and 15%, respectively. G1 + xerostomia was predicted reasonably well with area under the ROC curve ranging from 0.69 to 0.76. The LKB model failed to provide reliable G2 + xerostomia predictions at both time points.

CONCLUSIONS

Reduction of the mean dose to parotid glands below QUANTEC guidelines resulted in low G2 + xerostomia rates. In this dose domain, the mean dose models predicted G1 + xerostomia fairly well, however, failed to recognize patients at risk of G2 + xerostomia. There is a need for the development of more flexible models able to capture complexity of dose response in this dose regime.