Roses Have Thorns: Understanding the Downside of Oncological Care Delivery Through Visual Analytics and Sequential Rule Mining.

Personalized head and neck cancer therapeutics have greatly improved survival rates for patients, but are often leading to understudied long-lasting symptoms which affect quality of life. Sequential rule mining (SRM) is a promising unsupervised machine learning method for predicting longitudinal patterns in temporal data which, however, can output many repetitive patterns that are difficult to interpret without the assistance of visual analytics. We present a data-driven, human-machine analysis visual system developed in collaboration with SRM model builders in cancer symptom research, which facilitates mechanistic knowledge discovery in large scale, multivariate cohort symptom data. Our system supports multivariate predictive modeling of post-treatment symptoms based on during-treatment symptoms. It supports this goal through an SRM, clustering, and aggregation back end, and a custom front end to help develop and tune the predictive models. The system also explains the resulting predictions in the context of therapeutic decisions typical in personalized care delivery. We evaluate the resulting models and system with an interdisciplinary group of modelers and head and neck oncology researchers. The results demonstrate that our system effectively supports clinical and symptom research.

Multi-organ spatial stratification of 3-D dose distributions improves risk prediction of long-term self-reported severe symptoms in oropharyngeal cancer patients receiving radiotherapy: development of a pre-treatment decision support tool.

Purpose: Identify Oropharyngeal cancer (OPC) patients at high-risk of developing long-term severe radiation-associated symptoms using dose volume histograms for organs-at-risk, via unsupervised clustering. Material and methods: All patients were treated using radiation therapy for OPC. Dose-volume histograms of organs-at-risk were extracted from patients' treatment plans. Symptom ratings were collected via the MD Anderson Symptom Inventory (MDASI) given weekly during, and 6 months post-treatment. Drymouth, trouble swallowing, mucus, and vocal dysfunction were selected for analysis in this study. Patient stratifications were obtained by applying Bayesian Mixture Models with three components to patient's dose histograms for relevant organs. The clusters with the highest total mean doses were translated into dose thresholds using rule mining. Patient stratifications were compared against Tumor staging information using multivariate likelihood ratio tests. Model performance for prediction of moderate/severe symptoms at 6 months was compared against normal tissue complication probability (NTCP) models using cross-validation. Results: A total of 349 patients were included for long-term symptom prediction. High-risk clusters were significantly correlated with outcomes for severe late drymouth (p <.0001, OR = 2.94), swallow (p = .002, OR = 5.13), mucus (p = .001, OR = 3.18), and voice (p = .009, OR = 8.99). Simplified clusters were also correlated with late severe symptoms for drymouth (p <.001, OR = 2.77), swallow (p = .01, OR = 3.63), mucus (p = .01, OR = 2.37), and voice (p <.001, OR = 19.75). Proposed cluster stratifications show better performance than NTCP models for severe drymouth (AUC.598 vs.559, MCC.143 vs.062), swallow (AUC.631 vs.561, MCC.20 vs -.030), mucus (AUC.596 vs.492, MCC.164 vs -.041), and voice (AUC.681 vs.555, MCC.181 vs -.019). Simplified dose thresholds also show better performance than baseline models for predicting late severe ratings for all symptoms. Conclusion: Our results show that leveraging the 3-D dose histograms from radiation therapy plan improves stratification of patients according to their risk of experiencing long-term severe radiation associated symptoms, beyond existing NTPC models. Our rule-based method can approximate our stratifications with minimal loss of accuracy and can proactively identify risk factors for radiation-associated toxicity.

Spatially-aware clustering improves AJCC-8 risk stratification performance in oropharyngeal carcinomas.

OBJECTIVE: Evaluate the effectiveness of machine learning tools that incorporate spatial information such as disease location and lymph node metastatic patterns-of-spread, for prediction of survival and toxicity in HPV+ oropharyngeal cancer (OPC). MATERIALS & METHODS: 675 HPV+ OPC patients that were treated at MD Anderson Cancer Center between 2005 and 2013 with curative intent IMRT were retrospectively collected under IRB approval. Risk stratifications incorporating patient radiometric data and lymph node metastasis patterns via an anatomically-adjacent representation with hierarchical clustering were identified. These clusterings were combined into a 3-level patient stratification and included along with other known clinical features in a Cox model for predicting survival outcomes, and logistic regression for toxicity, using independent subsets for training and validation. RESULTS: Four groups were identified and combined into a 3-level stratification. The inclusion of patient stratifications in predictive models for 5-yr Overall survival (OS), 5-year recurrence free survival, (RFS) and Radiation-associated dysphagia (RAD) consistently improved model performance measured using the area under the curve (AUC). Test set AUC improvements over models with clinical covariates, was 9 % for predicting OS, and 18 % for predicting RFS, and 7 % for predicting RAD. For models with both clinical and AJCC covariates, AUC improvement was 7 %, 9 %, and 2 % for OS, RFS, and RAD, respectively. CONCLUSION: Including data-driven patient stratifications considerably improve prognosis for survival and toxicity outcomes over the performance achieved by clinical staging and clinical covariates alone. These stratifications generalize well to across cohorts, and sufficient information for reproducing these clusters is included.

Evaluating Autoencoders for Dimensionality Reduction of MRI-derived Radiomics and Classification of Malignant Brain Tumors.

Malignant brain tumors including parenchymal metastatic (MET) lesions, glioblastomas (GBM), and lymphomas (LYM) account for 29.7% of brain cancers. However, the characterization of these tumors from MRI imaging is difficult due to the similarity of their radiologically observed image features. Radiomics is the extraction of quantitative imaging features to characterize tumor intensity, shape, and texture. Applying machine learning over radiomic features could aid diagnostics by improving the classification of these common brain tumors. However, since the number of radiomic features is typically larger than the number of patients in the study, dimensionality reduction is needed to balance feature dimensionality and model complexity. Autoencoders are a form of unsupervised representation learning that can be used for dimensionality reduction. It is similar to PCA but uses a more complex and non-linear model to learn a compact latent space. In this work, we examine the effectiveness of autoencoders for dimensionality reduction on the radiomic feature space of multiparametric MRI images and the classification of malignant brain tumors: GBM, LYM, and MET. We further aim to address the class imbalances imposed by the rarity of lymphomas by examining different approaches to increase overall predictive performance through multiclass decomposition strategies.

Improving Prediction of Late Symptoms using LSTM and Patient-reported Outcomes for Head and Neck Cancer Patients.

Patient-Reported Outcomes (PRO) are collected directly from the patients using symptom questionnaires. In the case of head and neck cancer patients, PRO surveys are recorded every week during treatment with each patient's visit to the clinic and at different follow-up times after the treatment has concluded. PRO surveys can be very informative regarding the patient's status and the effect of treatment on the patient's quality of life (QoL). Processing PRO data is challenging for several reasons. First, missing data is frequent as patients might skip a question or a questionnaire altogether. Second, PROs are patient-dependent, a rating of 5 for one patient might be a rating of 10 for another patient. Finally, most patients experience severe symptoms during treatment which usually subside over time. However, for some patients, late toxicities persist negatively affecting the patient's QoL. These long-term severe symptoms are hard to predict and are the focus of this study. In this work, we model PRO data collected from head and neck cancer patients treated at the MD Anderson Cancer Center using the MD Anderson Symptom Inventory (MDASI) questionnaire as time series. We impute missing values with a combination of K nearest neighbor (KNN) and Long Short-Term Memory (LSTM) neural networks, and finally, apply LSTM to predict late symptom severity 12 months after treatment. We compare performance against clinical and ARIMA models. We show that the LSTM model combined with KNN imputation is effective in predicting late-stage symptom ratings for occurrence and severity under the AUC and F1 score metrics.

A Tale of Two Centers: Visual Exploration of Health Disparities in Cancer Care.

The annual incidence of head and neck cancers (HNC) worldwide is more than 550,000 cases, with around 300,000 deaths each year. However, the incidence rates and disease-characteristics of HNC differ between treatment centers and different populations, due to undetermined reasons, which may or not include socioeconomic factors. The multi-faceted and multi-variate nature of the data in the context of the emerging field of health disparities research makes automated analysis impractical. Hence, we present a visual analysis approach to explore the health disparities in the data of HNC patients from two different cohorts at two cancer care centers. Our approach integrates data from multiple sources, including census data and city data, with custom visual encodings and with a nearest neighbor approach. Our design, created in collaboration with oncology experts, makes it possible to analyze the patients' demographic, disease characteristics, treatments and outcomes, and to make significant comparisons of these two cohorts and of individual patients. We evaluate this approach through two case studies performed with domain experts. The results demonstrate that this visual analysis approach successfully accomplishes the goal of comparing two cohorts in terms of different significant factors, and can provide insights into the main source of health disparities between the two centers.

Optimal Treatment Selection in Sequential Systemic and Locoregional Therapy of Oropharyngeal Squamous Carcinomas: Deep Q-Learning With a Patient-Physician Digital Twin Dyad.

BACKGROUND: Currently, selection of patients for sequential versus concurrent chemotherapy and radiation regimens lacks evidentiary support and it is based on locally optimal decisions for each step. OBJECTIVE: We aim to optimize the multistep treatment of patients with head and neck cancer and predict multiple patient survival and toxicity outcomes, and we develop, apply, and evaluate a first application of deep Q-learning (DQL) and simulation to this problem. METHODS: The treatment decision DQL digital twin and the patient's digital twin were created, trained, and evaluated on a data set of 536 patients with oropharyngeal squamous cell carcinoma with the goal of, respectively, determining the optimal treatment decisions with respect to survival and toxicity metrics and predicting the outcomes of the optimal treatment on the patient. Of the data set of 536 patients, the models were trained on a subset of 402 (75%) patients (split randomly) and evaluated on a separate set of 134 (25%) patients. Training and evaluation of the digital twin dyad was completed in August 2020. The data set includes 3-step sequential treatment decisions and complete relevant history of the patient cohort treated at MD Anderson Cancer Center between 2005 and 2013, with radiomics analysis performed for the segmented primary tumor volumes. RESULTS: On the test set, we found mean 87.35% (SD 11.15%) and median 90.85% (IQR 13.56%) accuracies in treatment outcome prediction, matching the clinicians' outcomes and improving the (predicted) survival rate by +3.73% (95% CI -0.75% to 8.96%) and the dysphagia rate by +0.75% (95% CI -4.48% to 6.72%) when following DQL treatment decisions. CONCLUSIONS: Given the prediction accuracy and predicted improvement regarding the medically relevant outcomes yielded by this approach, this digital twin dyad of the patient-physician dynamic treatment problem has the potential of aiding physicians in determining the optimal course of treatment and in assessing its outcomes.

THALIS: Human-Machine Analysis of Longitudinal Symptoms in Cancer Therapy.

Although cancer patients survive years after oncologic therapy, they are plagued with long-lasting or permanent residual symptoms, whose severity, rate of development, and resolution after treatment vary largely between survivors. The analysis and interpretation of symptoms is complicated by their partial co-occurrence, variability across populations and across time, and, in the case of cancers that use radiotherapy, by further symptom dependency on the tumor location and prescribed treatment. We describe THALIS, an environment for visual analysis and knowledge discovery from cancer therapy symptom data, developed in close collaboration with oncology experts. Our approach leverages unsupervised machine learning methodology over cohorts of patients, and, in conjunction with custom visual encodings and interactions, provides context for new patients based on patients with similar diagnostic features and symptom evolution. We evaluate this approach on data collected from a cohort of head and neck cancer patients. Feedback from our clinician collaborators indicates that THALIS supports knowledge discovery beyond the limits of machines or humans alone, and that it serves as a valuable tool in both the clinic and symptom research.

Identifying Symptom Clusters Through Association Rule Mining.

Cancer patients experience many symptoms throughout their cancer treatment and sometimes suffer from lasting effects post-treatment. Patient-Reported Outcome (PRO) surveys provide a means for monitoring the patient's symptoms during and after treatment. Symptom cluster (SC) research seeks to understand these symptoms and their relationships to define new treatment and disease management methods to improve patient's quality of life. This paper introduces association rule mining (ARM) as a novel alternative for identifying symptom clusters. We compare the results to prior research and find that while some of the SCs are similar, ARM uncovers more nuanced relationships between symptoms such as anchor symptoms that serve as connections between interference and cancer-specific symptoms.

Oropharyngeal cancer patient stratification using random forest based-learning over high-dimensional radiomic features.

To improve risk prediction for oropharyngeal cancer (OPC) patients using cluster analysis on the radiomic features extracted from pre-treatment Computed Tomography (CT) scans. 553 OPC Patients randomly split into training (80%) and validation (20%), were classified into 2 or 3 risk groups by applying hierarchical clustering over the co-occurrence matrix obtained from a random survival forest (RSF) trained over 301 radiomic features. The cluster label was included together with other clinical data to train an ensemble model using five predictive models (Cox, random forest, RSF, logistic regression, and logistic-elastic net). Ensemble performance was evaluated over the independent test set for both recurrence free survival (RFS) and overall survival (OS). The Kaplan-Meier curves for OS stratified by cluster label show significant differences for both training and testing (p val < 0.0001). When compared to the models trained using clinical data only, the inclusion of the cluster label improves AUC test performance from .62 to .79 and from .66 to .80 for OS and RFS, respectively. The extraction of a single feature, namely a cluster label, to represent the high-dimensional radiomic feature space reduces the dimensionality and sparsity of the data. Moreover, inclusion of the cluster label improves model performance compared to clinical data only and offers comparable performance to the models including raw radiomic features.

Precision association of lymphatic disease spread with radiation-associated toxicity in oropharyngeal squamous carcinomas.

PURPOSE: To determine whether patient similarity in terms of head and neck cancer spread through lymph nodes correlates significantly with radiation-associated toxicity. MATERIALS AND METHODS: 582 head and neck cancer patients received radiotherapy for oropharyngeal cancer (OPC) and had non-metastatic affected lymph nodes in the head and neck. Affected lymph nodes were segmented from pretreatment contrast-enhanced tomography scans and categorized according to consensus guidelines. Similar patients were clustered into 4 groups according to a graph-based representation of disease spread through affected lymph nodes. Correlation between dysphagia-associated symptoms and patient groups was calculated. RESULTS: Out of 582 patients, 26% (152) experienced toxicity during a follow up evaluation 6 months after completion of radiotherapy treatment. Patient groups identified by our approach were significantly correlated with dysphagia, feeding tube, and aspiration toxicity (p < .0005). DISCUSSION: Our results suggest that structural geometry-aware characterization of affected lymph nodes can be used to better predict radiation-associated dysphagia at time of diagnosis, and better inform treatment guidelines. CONCLUSION: Our work successfully stratified a patient cohort into similar groups using a structural geometry, graph-encoding of affected lymph nodes in oropharyngeal cancer patients, that were predictive of late radiation-associated dysphagia and toxicity.

Predicting late symptoms of head and neck cancer treatment using LSTM and patient reported outcomes.

Patient-Reported Outcome (PRO) surveys are used to monitor patients' symptoms during and after cancer treatment. Acute symptoms refer to those experienced during treatment and late symptoms refer to those experienced after treatment. While most patients experience severe symptoms during treatment, these usually subside in the late stage. However, for some patients, late toxicities persist negatively affecting the patient's quality of life (QoL). In the case of head and neck cancer patients, PRO surveys are recorded every week during the patient's visit to the clinic and at different follow-up times after the treatment has concluded. In this paper, we model the PRO data as a time-series and apply Long-Short Term Memory (LSTM) neural networks for predicting symptom severity in the late stage. The PRO data used in this project corresponds to MD Anderson Symptom Inventory (MDASI) questionnaires collected from head and neck cancer patients treated at the MD Anderson Cancer Center. We show that the LSTM model is effective in predicting symptom ratings under the RMSE and NRMSE metrics. Our experiments show that the LSTM model also outperforms other machine learning models and time-series prediction models for these data.

Precision toxicity correlates of tumor spatial proximity to organs at risk in cancer patients receiving intensity-modulated radiotherapy.

PURPOSE: Using a 200 Head and Neck cancer (HNC) patient cohort, we employ patient similarity based on tumor location, volume, and proximity to organs at risk to predict radiation-associated dysphagia (RAD) in a new patient receiving intensity modulated radiation therapy (IMRT). MATERIAL AND METHODS: All patients were treated using curative-intent IMRT. Anatomical features were extracted from contrast-enhanced tomography scans acquired pre-treatment. Patient similarity was computed using a topological similarity measure, which allowed for the prediction of normal tissues' mean doses. We performed feature selection and clustering, and used the resulting groups of patients to forecast RAD. We used Logistic Regression (LG) cross-validation to assess the potential toxicity risk of these groupings. RESULTS: Out of 200 patients, 34 patients were recorded as having RAD. Patient clusters were significantly correlated with RAD (p < .0001). The area under the receiver-operator curve (AUC) using pre-established, baseline features gave a predictive accuracy of 0.79, while the addition of our cluster labels improved accuracy to 0.84. CONCLUSION: Our results show that spatial information available pre-treatment can be used to robustly identify groups of RAD high-risk patients. We identify feature sets that considerably improve toxicity risk prediction beyond what is possible using baseline features. Our results also suggest that similarity-based predicted mean doses to organs can be used as valid predictors of risk to organs.

Clustering of Largely Right-Censored Oropharyngeal Head and Neck Cancer Patients for Discriminative Groupings to Improve Outcome Prediction.

Clustering is the task of identifying groups of similar subjects according to certain criteria. The AJCC staging system can be thought as a clustering mechanism that groups patients based on their disease stage. This grouping drives prognosis and influences treatment. The goal of this work is to evaluate the efficacy of machine learning algorithms to cluster the patients into discriminative groups to improve prognosis for overall survival (OS) and relapse free survival (RFS) outcomes. We apply clustering over a retrospectively collected data from 644 head and neck cancer patients including both clinical and radiomic features. In order to incorporate outcome information into the clustering process and deal with the large proportion of censored samples, the feature space was scaled using the regression coefficients fitted using a proxy dependent variable, martingale residuals, instead of follow-up time. Two clusters were identified and evaluated using cross validation. The Kaplan Meier (KM) curves between the two clusters differ significantly for OS and RFS (p-value < 0.0001). Moreover, there was a relative predictive improvement when using the cluster label in addition to the clinical features compared to using only clinical features where AUC increased by 5.7% and 13.0% for OS and RFS, respectively.

Precision Risk Analysis of Cancer Therapy with Interactive Nomograms and Survival Plots.

We present the design and evaluation of an integrated problem solving environment for cancer therapy analysis. The environment intertwines a statistical martingale model and a K Nearest Neighbor approach with visual encodings, including novel interactive nomograms, in order to compute and explain a patient's probability of survival as a function of similar patient results. A coordinated views paradigm enables exploration of the multivariate, heterogeneous and few-valued data from a large head and neck cancer repository. A visual scaffolding approach further enables users to build from familiar representations to unfamiliar ones. Evaluation with domain experts show how this visualization approach and set of streamlined workflows enable the systematic and precise analysis of a patient prognosis in the context of cohorts of similar patients. We describe the design lessons learned from this successful, multi-site remote collaboration.

Machine Learning Applications in Head and Neck Radiation Oncology: Lessons From Open-Source Radiomics Challenges.

Radiomics leverages existing image datasets to provide non-visible data extraction via image post-processing, with the aim of identifying prognostic, and predictive imaging features at a sub-region of interest level. However, the application of radiomics is hampered by several challenges such as lack of image acquisition/analysis method standardization, impeding generalizability. As of yet, radiomics remains intriguing, but not clinically validated. We aimed to test the feasibility of a non-custom-constructed platform for disseminating existing large, standardized databases across institutions for promoting radiomics studies. Hence, University of Texas MD Anderson Cancer Center organized two public radiomics challenges in head and neck radiation oncology domain. This was done in conjunction with MICCAI 2016 satellite symposium using Kaggle-in-Class, a machine-learning and predictive analytics platform. We drew on clinical data matched to radiomics data derived from diagnostic contrast-enhanced computed tomography (CECT) images in a dataset of 315 patients with oropharyngeal cancer. Contestants were tasked to develop models for (i) classifying patients according to their human papillomavirus status, or (ii) predicting local tumor recurrence, following radiotherapy. Data were split into training, and test sets. Seventeen teams from various professional domains participated in one or both of the challenges. This review paper was based on the contestants' feedback; provided by 8 contestants only (47%). Six contestants (75%) incorporated extracted radiomics features into their predictive model building, either alone (n = 5; 62.5%), as was the case with the winner of the "HPV" challenge, or in conjunction with matched clinical attributes (n = 2; 25%). Only 23% of contestants, notably, including the winner of the "local recurrence" challenge, built their model relying solely on clinical data. In addition to the value of the integration of machine learning into clinical decision-making, our experience sheds light on challenges in sharing and directing existing datasets toward clinical applications of radiomics, including hyper-dimensionality of the clinical/imaging data attributes. Our experience may help guide researchers to create a framework for sharing and reuse of already published data that we believe will ultimately accelerate the pace of clinical applications of radiomics; both in challenge or clinical settings.

A prospective in silico analysis of interdisciplinary and interobserver spatial variability in post-operative target delineation of high-risk oral cavity cancers: Does physician specialty matter?

BACKGROUND: The aim of this study was to determine the interdisciplinary agreement in identifying the post-operative tumor bed. METHODS: Three radiation oncologists (ROs), four surgeons, and three radiologists segmented post-operative tumor and nodal beds for three patients with oral cavity cancer. Specialty cohort composite contours were created by STAPLE algorithm implementation results for interspecialty comparison. Dice similarity coefficient and Hausdorff distance were utilized to compare spatial differentials between specialties. RESULTS: There were significant differences between disciplines in target delineation. There was unacceptable variation in Dice similarity coefficient for each observer and discipline when compared to the STAPLE contours. Within surgery and radiology disciplines, there was good consistency in volumes. ROs and radiologists have similar Dice similarity coefficient scores compared to surgeons. CONCLUSION: There were significant interdisciplinary differences in perceptions of tissue-at-risk. Better communication and explicit description of at-risk areas between disciplines is required to ensure high-risk areas are adequately targeted.

Evaluating the Effect of Right-Censored End Point Transformation for Radiomic Feature Selection of Data From Patients With Oropharyngeal Cancer.

PURPOSE: To evaluate the effect of transforming a right-censored outcome into binary, continuous, and censored-aware representations on radiomics feature selection and subsequent prediction of overall survival (OS) and relapse-free survival (RFS) of patients with oropharyngeal cancer. METHODS AND MATERIALS: Different feature selection techniques were applied using a binary outcome indicating event occurrence before median follow-up time, a continuous outcome using the Martingale residuals from a proportional hazards model, and the raw right-censored time-to-event outcome. Radiomic signatures combined with clinical variables were used for risk prediction. Three metrics for accuracy and calibration were used to evaluate eight feature selectors and six predictive models. RESULTS: Feature selection across 529 patients on more than 3,800 radiomic features resulted in increases ranging from 0.01 to 0.11 in C-index and area under the curve (AUC) scores compared with clinical features alone. The ensemble model yielded the best scores for AUC and C-index (often > 0.7) and calibration (Nam-D'Agostino test statistic often < 15.5 with 8 df). The random forest feature selectors achieved the best performance considering all metrics. Random regression forest performed the best in OS prediction with the ensemble model (AUC, 0.75; C-index, 0.76; calibration, 8.7). Random survival forest performed the best in RFS prediction with the ensemble model (AUC, 0.71; C-index, 0.68; calibration, 19.1). CONCLUSION: Including a radiomic signature results in better prediction than using only clinical data. Signatures generated randomly or without considering the outcome result in poor calibration scores. The random forest feature selectors for each of the three transformations typically selected the greatest number of features and produced the best predictions at acceptable calibration levels. In particular, random regression forest and random survival forest performed best for OS and RFS, respectively.