A Novel Approach to Characterize State-level Food Environment and Predict Obesity Rate Using Social Media Data: Correlational Study.

BACKGROUND: Community obesity outcomes can reflect the food environment to which the community belongs. Recent studies have suggested that the local food environment can be measured by the degree of food accessibility, and survey data are normally used to calculate food accessibility. However, compared with survey data, social media data are organic, continuously updated, and cheaper to collect. OBJECTIVE: The objective of our study was to use publicly available social media data to learn the relationship between food environment and obesity rates at the state level. METHODS: To characterize the caloric information of the local food environment, we used food categories from Yelp and collected caloric information from MyFitnessPal for each category based on their popular dishes. We then calculated the average calories for each category and created a weighted score for each state. We also calculated 2 other dimensions from the concept of access, acceptability and affordability, to build obesity prediction models. RESULTS: The local food environment characterized using only publicly available social media data had a statistically significant correlation with the state obesity rate. We achieved a Pearson correlation of 0.796 between the predicted obesity rate and the reported obesity rate from the Behavioral Risk Factor Surveillance System across US states and the District of Columbia. The model with 3 generated feature sets achieved the best performance. CONCLUSIONS: Our study proposed a method for characterizing state-level food environments only using continuously updated social media data. State-level food environments were accurately described using social media data, and the model also showed a disparity in the available food between states with different obesity rates. The proposed method should elastically apply to local food environments of different sizes and predict obesity rates effectively.

A Novel Machine Learning Framework for Comparison of Viral COVID-19-Related Sina Weibo and Twitter Posts: Workflow Development and Content Analysis.

BACKGROUND: Social media plays a critical role in health communications, especially during global health emergencies such as the current COVID-19 pandemic. However, there is a lack of a universal analytical framework to extract, quantify, and compare content features in public discourse of emerging health issues on different social media platforms across a broad sociocultural spectrum. OBJECTIVE: We aimed to develop a novel and universal content feature extraction and analytical framework and contrast how content features differ with sociocultural background in discussions of the emerging COVID-19 global health crisis on major social media platforms. METHODS: We sampled the 1000 most shared viral Twitter and Sina Weibo posts regarding COVID-19, developed a comprehensive coding scheme to identify 77 potential features across six major categories (eg, clinical and epidemiological, countermeasures, politics and policy, responses), quantified feature values (0 or 1, indicating whether or not the content feature is mentioned in the post) in each viral post across social media platforms, and performed subsequent comparative analyses. Machine learning dimension reduction and clustering analysis were then applied to harness the power of social media data and provide more unbiased characterization of web-based health communications. RESULTS: There were substantially different distributions, prevalence, and associations of content features in public discourse about the COVID-19 pandemic on the two social media platforms. Weibo users were more likely to focus on the disease itself and health aspects, while Twitter users engaged more about policy, politics, and other societal issues. CONCLUSIONS: We extracted a rich set of content features from social media data to accurately characterize public discourse related to COVID-19 in different sociocultural backgrounds. In addition, this universal framework can be adopted to analyze social media discussions of other emerging health issues beyond the COVID-19 pandemic.

A Multimodality Machine Learning Approach to Differentiate Severe and Nonsevere COVID-19: Model Development and Validation.

BACKGROUND: Effectively and efficiently diagnosing patients who have COVID-19 with the accurate clinical type of the disease is essential to achieve optimal outcomes for the patients as well as to reduce the risk of overloading the health care system. Currently, severe and nonsevere COVID-19 types are differentiated by only a few features, which do not comprehensively characterize the complicated pathological, physiological, and immunological responses to SARS-CoV-2 infection in the different disease types. In addition, these type-defining features may not be readily testable at the time of diagnosis. OBJECTIVE: In this study, we aimed to use a machine learning approach to understand COVID-19 more comprehensively, accurately differentiate severe and nonsevere COVID-19 clinical types based on multiple medical features, and provide reliable predictions of the clinical type of the disease. METHODS: For this study, we recruited 214 confirmed patients with nonsevere COVID-19 and 148 patients with severe COVID-19. The clinical characteristics (26 features) and laboratory test results (26 features) upon admission were acquired as two input modalities. Exploratory analyses demonstrated that these features differed substantially between two clinical types. Machine learning random forest models based on all the features in each modality as well as on the top 5 features in each modality combined were developed and validated to differentiate COVID-19 clinical types. RESULTS: Using clinical and laboratory results independently as input, the random forest models achieved >90% and >95% predictive accuracy, respectively. The importance scores of the input features were further evaluated, and the top 5 features from each modality were identified (age, hypertension, cardiovascular disease, gender, and diabetes for the clinical features modality, and dimerized plasmin fragment D, high sensitivity troponin I, absolute neutrophil count, interleukin 6, and lactate dehydrogenase for the laboratory testing modality, in descending order). Using these top 10 multimodal features as the only input instead of all 52 features combined, the random forest model was able to achieve 97% predictive accuracy. CONCLUSIONS: Our findings shed light on how the human body reacts to SARS-CoV-2 infection as a unit and provide insights on effectively evaluating the disease severity of patients with COVID-19 based on more common medical features when gold standard features are not available. We suggest that clinical information can be used as an initial screening tool for self-evaluation and triage, while laboratory test results should be applied when accuracy is the priority.

Accurately Differentiating Between Patients With COVID-19, Patients With Other Viral Infections, and Healthy Individuals: Multimodal Late Fusion Learning Approach.

BACKGROUND: Effectively identifying patients with COVID-19 using nonpolymerase chain reaction biomedical data is critical for achieving optimal clinical outcomes. Currently, there is a lack of comprehensive understanding in various biomedical features and appropriate analytical approaches for enabling the early detection and effective diagnosis of patients with COVID-19. OBJECTIVE: We aimed to combine low-dimensional clinical and lab testing data, as well as high-dimensional computed tomography (CT) imaging data, to accurately differentiate between healthy individuals, patients with COVID-19, and patients with non-COVID viral pneumonia, especially at the early stage of infection. METHODS: In this study, we recruited 214 patients with nonsevere COVID-19, 148 patients with severe COVID-19, 198 noninfected healthy participants, and 129 patients with non-COVID viral pneumonia. The participants' clinical information (ie, 23 features), lab testing results (ie, 10 features), and CT scans upon admission were acquired and used as 3 input feature modalities. To enable the late fusion of multimodal features, we constructed a deep learning model to extract a 10-feature high-level representation of CT scans. We then developed 3 machine learning models (ie, k-nearest neighbor, random forest, and support vector machine models) based on the combined 43 features from all 3 modalities to differentiate between the following 4 classes: nonsevere, severe, healthy, and viral pneumonia. RESULTS: Multimodal features provided substantial performance gain from the use of any single feature modality. All 3 machine learning models had high overall prediction accuracy (95.4%-97.7%) and high class-specific prediction accuracy (90.6%-99.9%). CONCLUSIONS: Compared to the existing binary classification benchmarks that are often focused on single-feature modality, this study's hybrid deep learning-machine learning framework provided a novel and effective breakthrough for clinical applications. Our findings, which come from a relatively large sample size, and analytical workflow will supplement and assist with clinical decision support for current COVID-19 diagnostic methods and other clinical applications with high-dimensional multimodal biomedical features.

[Contrasting Misinformation and Real-Information Dissemination Network Structures on Social Media During a Health Emergency]. / Comparación de las estructuras de difusión de información errónea y verídica en las redes sociales durante una emergencia de salud pública.

OBJECTIVES: To provide a comprehensive workflow to identify top influential health misinformation about Zika on Twitter in 2016, reconstruct information dissemination networks of retweeting, contrast mis- from real information on various metrics, and investigate how Zika misinformation proliferated on social media during the Zika epidemic. METHODS: We systematically reviewed the top 5000 English-language Zika tweets, established an evidence-based definition of "misinformation," identified misinformation tweets, and matched a comparable group of real-information tweets. We developed an algorithm to reconstruct retweeting networks for 266 misinformation and 458 comparable real-information tweets. We computed and compared 9 network metrics characterizing network structure across various levels between the 2 groups. RESULTS: There were statistically significant differences in all 9 network metrics between real and misinformation groups. Misinformation network structures were generally more sophisticated than those in the real-information group. There was substantial within-group variability, too. CONCLUSIONS: Dissemination networks of Zika misinformation differed substantially from real information on Twitter, indicating that misinformation utilized distinct dissemination mechanisms from real information. Our study will lead to a more holistic understanding of health misinformation challenges on social media.

Contrasting Misinformation and Real-Information Dissemination Network Structures on Social Media During a Health Emergency.

Objectives. To provide a comprehensive workflow to identify top influential health misinformation about Zika on Twitter in 2016, reconstruct information dissemination networks of retweeting, contrast mis- from real information on various metrics, and investigate how Zika misinformation proliferated on social media during the Zika epidemic.Methods. We systematically reviewed the top 5000 English-language Zika tweets, established an evidence-based definition of "misinformation," identified misinformation tweets, and matched a comparable group of real-information tweets. We developed an algorithm to reconstruct retweeting networks for 266 misinformation and 458 comparable real-information tweets. We computed and compared 9 network metrics characterizing network structure across various levels between the 2 groups.Results. There were statistically significant differences in all 9 network metrics between real and misinformation groups. Misinformation network structures were generally more sophisticated than those in the real-information group. There was substantial within-group variability, too.Conclusions. Dissemination networks of Zika misinformation differed substantially from real information on Twitter, indicating that misinformation utilized distinct dissemination mechanisms from real information. Our study will lead to a more holistic understanding of health misinformation challenges on social media.

Quantitative comparison of automatic and manual IMRT optimization for prostate cancer: the benefits of DVH prediction.

A recent publication indicated that the patient anatomical feature (PAF) model was capable of predicting optimal objectives based on past experience. In this study, the benefits of IMRT optimization using PAF-predicted objectives as guidance for prostate were evaluated. Three different optimization methods were compared.1) Expert Plan: Ten prostate cases (16 plans) were planned by an expert planner using conventional trial-and-error approach started with institutional modified OAR and PTV constraints. Optimization was stopped at 150 iterations and that plan was saved as Expert Plan. 2) Clinical Plan: The planner would keep working on the Expert Plan till he was satisfied with the dosimetric quality and the final plan was referred to as Clinical Plan. 3) PAF Plan: A third sets of plans for the same ten patients were generated fully automatically using predicted DVHs as guidance. The optimization was based on PAF-based predicted objectives, and was continued to 150 iterations without human interaction. DMAX and D98% for PTV, DMAX for femoral heads, DMAX, D10cc, D25%/D17%, and D40% for bladder/rectum were compared. Clinical Plans are further optimized with more iterations and adjustments, but in general provided limited dosimetric benefits over Expert Plans. PTV D98% agreed within 2.31% among Expert, Clinical, and PAF plans. Between Clinical and PAF Plans, differences for DMAX of PTV, bladder, and rectum were within 2.65%, 2.46%, and 2.20%, respectively. Bladder D10cc was higher for PAF but < 1.54% in general. Bladder D25% and D40% were lower for PAF, by up to 7.71% and 6.81%, respectively. Rectum D10cc, D17%, and D40% were 2.11%, 2.72%, and 0.27% lower for PAF, respectively. DMAX for femoral heads were comparable (< 35 Gy on average). Compared to Clinical Plan (Primary + Boost), the average optimization time for PAF plan was reduced by 5.2 min on average, with a maximum reduction of 7.1min. Total numbers of MUs per plan for PAF Plans were lower than Clinical Plans, indicating better delivery efficiency. The PAF-guided planning process is capable of generating clinical-quality prostate IMRT plans with no human intervention. Compared to manual optimization, this automatic optimization increases planning and delivery efficiency, while maintainingplan quality.

Dynamic Associations Between Centers for Disease Control and Prevention Social Media Contents and Epidemic Measures During COVID-19: Infoveillance Study.

BACKGROUND: Health agencies have been widely adopting social media to disseminate important information, educate the public on emerging health issues, and understand public opinions. The Centers for Disease Control and Prevention (CDC) widely used social media platforms during the COVID-19 pandemic to communicate with the public and mitigate the disease in the United States. It is crucial to understand the relationships between the CDC's social media communications and the actual epidemic metrics to improve public health agencies' communication strategies during health emergencies. OBJECTIVE: This study aimed to identify key topics in tweets posted by the CDC during the pandemic, investigate the temporal dynamics between these key topics and the actual COVID-19 epidemic measures, and make recommendations for the CDC's digital health communication strategies for future health emergencies. METHODS: Two types of data were collected: (1) a total of 17,524 COVID-19-related English tweets posted by the CDC between December 7, 2019, and January 15, 2022, and (2) COVID-19 epidemic measures in the United States from the public GitHub repository of Johns Hopkins University from January 2020 to July 2022. Latent Dirichlet allocation topic modeling was applied to identify key topics from all COVID-19-related tweets posted by the CDC, and the final topics were determined by domain experts. Various multivariate time series analysis techniques were applied between each of the identified key topics and actual COVID-19 epidemic measures to quantify the dynamic associations between these 2 types of time series data. RESULTS: Four major topics from the CDC's COVID-19 tweets were identified: (1) information on the prevention of health outcomes of COVID-19; (2) pediatric intervention and family safety; (3) updates of the epidemic situation of COVID-19; and (4) research and community engagement to curb COVID-19. Multivariate analyses showed that there were significant variabilities of progression between the CDC's topics and the actual COVID-19 epidemic measures. Some CDC topics showed substantial associations with the COVID-19 measures over different time spans throughout the pandemic, expressing similar temporal dynamics between these 2 types of time series data. CONCLUSIONS: Our study is the first to comprehensively investigate the dynamic associations between topics discussed by the CDC on Twitter and the COVID-19 epidemic measures in the United States. We identified 4 major topic themes via topic modeling and explored how each of these topics was associated with each major epidemic measure by performing various multivariate time series analyses. We recommend that it is critical for public health agencies, such as the CDC, to update and disseminate timely and accurate information to the public and align major topics with key epidemic measures over time. We suggest that social media can help public health agencies to inform the public on health emergencies and to mitigate them effectively.

Cerebral white matter hyperintensity in African Americans and European Americans with type 2 diabetes.

Previous studies involving inner city populations detected higher cerebral white matter hyperintensity (WMH) scores in African Americans (AAs) compared with European Americans (EAs). This finding might be attributable to the higher prevalence of cardiovascular disease (CVD) risk factors and poorer access to healthcare in AAs. Despite racial differences in CVD risk factor profiles, AAs have paradoxically lower levels of subclinical CVD. We hypothesized that AAs with diabetes and good access to healthcare would have comparable or lower levels of WMH as EAs. Racial differences in the distribution of WMH were analyzed in 46 AAs and 156 EAs with type 2 diabetes enrolled in the Diabetes Heart Study (DHS)-Mind, and replicated in a sample of 113 AAs and 61 EAs patients who had clinically indicated cerebral magnetic resonance imaging. Wilcoxon 2-sample tests and linear models were used to compare the distribution of WMH in AAs and EAs and to test for association between WMH and race. The unadjusted mean WMH score from the Diabetes Heart Study-Mind was 1.9 in AAs and 2.3 in EAs (P = .3244). Among those with clinically indicated magnetic resonance imaging, the mean WMH score was 2.9 in AAs and 3.9 in EAs (P = .0503). Adjustment for age and sex produced no statistically significant differences in WMH score between AAs and EAs. These independent datasets reveal comparable WMH scores in AAs and EAs, suggesting that disparities in access to healthcare and environmental exposures likely underlie the previously reported excess burden of WMH in AAs.

Content and sentiment surveillance (CSI): A critical component for modeling modern epidemics.

Comprehensive surveillance systems are the key to provide accurate data for effective modeling. Traditional symptom-based case surveillance has been joined with recent genomic, serologic, and environment surveillance to provide more integrated disease surveillance systems. A major gap in comprehensive disease surveillance is to accurately monitor potential population behavioral changes in real-time. Population-wide behaviors such as compliance with various interventions and vaccination acceptance significantly influence and drive the overall epidemic dynamics in the society. Original infoveillance utilizes online query data (e.g., Google and Wikipedia search of a specific content topic such as an epidemic) and later focuses on large volumes of online discourse data about the from social media platforms and further augments epidemic modeling. It mainly uses number of posts to approximate public awareness of the disease, and further compares with observed epidemic dynamics for better projection. The current COVID-19 pandemic shows that there is an urgency to further harness the rich, detailed content and sentiment information, which can provide more accurate and granular information on public awareness and perceptions toward multiple aspects of the disease, especially various interventions. In this perspective paper, we describe a novel conceptual analytical framework of content and sentiment infoveillance (CSI) and integration with epidemic modeling. This CSI framework includes data retrieval and pre-processing; information extraction via natural language processing to identify and quantify detailed time, location, content, and sentiment information; and integrating infoveillance with common epidemic modeling techniques of both mechanistic and data-driven methods. CSI complements and significantly enhances current epidemic models for more informed decision by integrating behavioral aspects from detailed, instantaneous infoveillance from massive social media data.

Health Information on Pre-Exposure Prophylaxis From Search Engines and Twitter: Readability Analysis.

BACKGROUND: Pre-exposure prophylaxis (PrEP) is proven to prevent HIV infection. However, PrEP uptake to date has been limited and inequitable. Analyzing the readability of existing PrEP-related information is important to understand the potential impact of available PrEP information on PrEP uptake and identify opportunities to improve PrEP-related education and communication. OBJECTIVE: We examined the readability of web-based PrEP information identified using search engines and on Twitter. We investigated the readability of web-based PrEP documents, stratified by how the PrEP document was obtained on the web, information source, document format and communication method, PrEP modality, and intended audience. METHODS: Web-based PrEP information in English was systematically identified using search engines and the Twitter API. We manually verified and categorized results and described the method used to obtain information, information source, document format and communication method, PrEP modality, and intended audience. Documents were converted to plain text for the analysis and readability of the collected documents was assessed using 4 readability indices. We conducted pairwise comparisons of readability based on how the PrEP document was obtained on the web, information source, document format, communication method, PrEP modality, and intended audience, then adjusted for multiple comparisons. RESULTS: A total of 463 documents were identified. Overall, the readability of web-based PrEP information was at a higher level (10.2-grade reading level) than what is recommended for health information provided to the general public (ninth-grade reading level, as suggested by the Department of Health and Human Services). Brochures (n=33, 7% of all identified resources) were the only type of PrEP materials that achieved the target of ninth-grade reading level. CONCLUSIONS: Web-based PrEP information is often written at a complex level for potential and current PrEP users to understand. This may hinder PrEP uptake for some people who would benefit from it. The readability of PrEP-related information found on the web should be improved to align more closely with health communication guidelines for reading level to improve access to this important health information, facilitate informed decisions by those with a need for PrEP, and realize national prevention goals for PrEP uptake and reducing new HIV infections in the United States.

Quantitative analysis of the factors which affect the interpatient organ-at-risk dose sparing variation in IMRT plans.

PURPOSE: The authors present an evidence-based approach to quantify the effects of an array of patient anatomical features of the planning target volumes (PTVs) and organs-at-risk (OARs) and their spatial relationships on the interpatient OAR dose sparing variation in intensity modulated radiation therapy (IMRT) plans by learning from a database of high-quality prior plans. METHODS: The authors formulized the dependence of OAR dose volume histograms (DVHs) on patient anatomical factors into feature models which were learned from prior plans by a stepwise multiple regression method. IMRT plans for 64 prostate, 82 head-and-neck (HN) treatments were used to train the models. Two major groups of anatomical features were considered in this study: the volumetric information and the spatial information. The geometry of OARs relative to PTV is represented by the distance-to-target histogram, DTH. Important anatomical and dosimetric features were extracted from DTH and DVH by principal component analysis. The final models were tested by additional 24 prostate and 24 HN plans. RESULTS: Significant patient anatomical factors contributing to OAR dose sparing in prostate and HN IMRT plans have been analyzed and identified. They are: the median distance between OAR and PTV, the portion of OAR volume within an OAR specific distance range, and the volumetric factors: the fraction of OAR volume which overlaps with PTV and the portion of OAR volume outside the primary treatment field. Overall, the determination coefficients R(2) for predicting the first principal component score (PCS1) of the OAR DVH by the above factors are above 0.68 for all the OARs and they are more than 0.53 for predicting the second principal component score (PCS2) of the OAR DVHs except brainstem and spinal cord. Thus, the above set of anatomical features combined has captured significant portions of the DVH variations for the OARs in prostate and HN plans. To test how well these features capture the interpatient organ dose sparing variations in general, the DVHs and specific dose-volume indices calculated from the regression models were compared with the actual DVHs and dose-volume indices from each patient's plan in the validation dataset. The dose-volume indices compared were V99%, V85%, and V50% for bladder and rectum in prostate plans and parotids median dose in HN plans. The authors found that for the bladder and rectum models, 17 out of 24 plans (71%) were within 6% OAR volume error and 21 plans (85%) were within 10% error; For the parotids model, the median dose values for 30 parotids out of 48 (63%) were within 6% prescription dose error and the values in 40 parotids (83%) were within 10% error. CONCLUSIONS: Quantitative analysis of patient anatomical features and their correlation with OAR dose sparing has identified a number of important factors that explain significant amount of interpatient DVH variations in OARs. These factors can be incorporated into evidence-based learning models as effective features to provide patient-specific OAR dose sparing goals.

Using Social Media to Understand Web-Based Social Factors Concerning Obesity: Systematic Review.

BACKGROUND: Evidence in the literature surrounding obesity suggests that social factors play a substantial role in the spread of obesity. Although social ties with a friend who is obese increase the probability of becoming obese, the role of social media in this dynamic remains underexplored in obesity research. Given the rapid proliferation of social media in recent years, individuals socialize through social media and share their health-related daily routines, including dieting and exercising. Thus, it is timely and imperative to review previous studies focused on social factors in social media and obesity. OBJECTIVE: This study aims to examine web-based social factors in relation to obesity research. METHODS: We conducted a systematic review. We searched PubMed, Association for Computing Machinery, and ScienceDirect for articles published by July 5, 2019. Web-based social factors that are related to obesity behaviors were studied and analyzed. RESULTS: In total, 1608 studies were identified from the selected databases. Of these 1608 studies, 50 (3.11%) studies met the eligibility criteria. In total, 10 types of web-based social factors were identified, and a socioecological model was adopted to explain their potential impact on an individual from varying levels of web-based social structure to social media users' connection to the real world. CONCLUSIONS: We found 4 levels of interaction in social media. Gender was the only factor found at the individual level, and it affects user's web-based obesity-related behaviors. Social support was the predominant factor identified, which benefits users in their weight loss journey at the interpersonal level. Some factors, such as stigma were also found to be associated with a healthy web-based social environment. Understanding the effectiveness of these factors is essential to help users create and maintain a healthy lifestyle.

Predictors of underutilization of lung cancer screening: a machine learning approach.

Lung cancer is the second common cancer and a leading cause of cancer-related death in the US. Unfavorably, the prevalence of using low-dose computed tomography (LDCT) for lung cancer prevention in the US has remained below 4% over time. The purpose of this study is to develop machine learning models to analyze interactive pathways of factors associated with lung cancer screening use with the LDCT. The study was based on the data retrieved from the 2018 Behavioral Risk Factor Surveillance System. After dealing with missing values, 86 variables and 710 samples were included in the decision tree model and the random forest model. The data were randomly split into training (569/710, 80%) and testing (141/710, 20%) sets. Gini impurity is used to select and determine the optimal split of the nodes in the model. Machine learning performance was evaluated by model accuracy, sensitivity, specificity, F1 score, etc. The average performance metrics of the decision tree model were obtained: average accuracy is 67.78%, F1 score is 65.76%, sensitivity is 62.52%, and specificity is 73.57% based on 100 runs. In the decision model, nine interactive pathways were identified among the following factors: average drinks per month, BMI, diabetes, first smoke age, years of smoking, year(s) quit smoking, sex, last sigmoidoscopy or colonoscopy, last dental visit, general health, insurance, education, and last Pap test. Lung cancer screening utilization is the result of the interplay of multifactors. Lung cancer screening programs in clinical settings should not only focus on patients' smoking behaviors but also consider other socioeconomic factors.

Technical note: Determining the applicability of a clinical knowledge-based learning model via prospective outlier detection.

BACKGROUND: Knowledge-based planning (KBP) is increasingly implemented clinically because of its demonstrated ability to improve treatment planning efficiency and reduce plan quality variations. However, cases with large dose-volume histogram (DVH) prediction uncertainties may still need manual adjustments by the planner to achieve high plan quality. PURPOSE: The purpose of this study is to develop a data-driven method to detect patients with high prediction uncertainties so that intentional effort is directed to these patients. METHODS: We apply an anomaly detection method known as the local outlier factor (LOF) to a dataset consisting of the training set and each of the prospective patients considered, to evaluate their likelihood of being an anomaly when compared with the training cases. Features used in the LOF analysis include anatomical features and the model-generated DVH principal component scores. To test the efficacy of the proposed model, 142 prostate patients were retrieved from the clinical database and split into a training dataset of 100 patients and a test dataset of 42 patients. The outlier identification performance was quantified by the difference between the DVH prediction root-mean-squared errors (RMSE) of the identified outlier cases and that of the remaining inlier cases. RESULTS: With a predefined LOF threshold of 1.4, the inlier cases achieved average RMSEs of 5.0 and 6.7 for bladder and rectum, while the outlier cases have substantially higher RMSEs of 6.7 and 13.0 in comparison. CONCLUSIONS: We propose a method that can determine the prospective patient's outlier status. This method can be integrated into existing automated treatment planning workflows to reduce the risk of generating suboptimal treatment plans while providing an upfront alert to the treatment planner.

Input feature design and its impact on the performance of deep learning models for predicting fluence maps in intensity-modulated radiation therapy.

Objective. Deep learning (DL) models for fluence map prediction (FMP) have great potential to reduce treatment planning time in intensity-modulated radiation therapy (IMRT) by avoiding the lengthy inverse optimization process. This study aims to improve the rigor of input feature design in a DL-FMP model by examining how different designs of input features influence model prediction performance.Approach. This study included 231 head-and-neck intensity-modulated radiation therapy patients. Three input feature designs were investigated. The first design (D1) assumed that information of all critical structures from all beam angles should be combined to predict fluence maps. The second design (D2) assumed that local anatomical information was sufficient for predicting radiation intensity of a beamlet at a respective beam angle. The third design (D3) assumed the need for both local anatomical information and inter-beam modulation to predict radiation intensity values of the beamlets that intersect at a voxel. For each input design, we tailored the DL model accordingly. All models were trained using the same set of ground truth plans (GT plans). The plans generated by DL models (DL plans) were analyzed using key dose-volume metrics. One-way ANOVA with multiple comparisons correction (Bonferroni method) was performed (significance level = 0.05).Main results. For PTV-related metrics, all DL plans had significantly higher maximum dose (p < 0.001), conformity index (p < 0.001), and heterogeneity index (p < 0.001) compared to GT plans, with D2 being the worst performer. Meanwhile, except for cord+5 mm (p < 0.001), DL plans of all designs resulted in OAR dose metrics that are comparable to those of GT plans.Significance. Local anatomical information contains most of the information that DL models need to predict fluence maps for clinically acceptable OAR sparing. Input features from beam angles are needed to achieve the best PTV coverage. These results provide valuable insights for further improvement of DL-FMP models and DL models in general.

Statins and Left Ventricular Ejection Fraction Following Doxorubicin Treatment.

BACKGROUND: Statins taken for cardiovascular indications by patients with breast cancer and lymphoma during doxorubicin treatment may attenuate left ventricular ejection fraction (LVEF) decline, but the effect of statins on LVEF among patients with no cardiovascular indications is unknown. METHODS: A double-blind, placebo-controlled, 24-month randomized trial of 40 mg of atorvastatin per day administered to patients with breast cancer and lymphoma receiving doxorubicin was conducted within the National Cancer Institute Community Oncology Research Program across 31 sites in the United States. At pretreatment and then 6 and 24 months after initiating doxorubicin, we assessed left ventricular (LV) volumes, strain, mass, and LVEF through cardiac magnetic resonance imaging, along with cognitive function and serum markers of inflammation. The primary outcome was the difference in 24-month LVEF between placebo and treatment groups, adjusted for pretreatment LVEF. RESULTS: A total of 279 participants were enrolled in the trial. Participants had a mean (±SD) age of 49±12 years; 92% were women; and 83% were White. The mean (±SD) LVEF values were 61.7±5.5% before treatment and 57.4±6.8% at 24 months in the placebo group and 62.6±6.4% before treatment and 57.7±5.6% at 24 months in the atorvastatin group. On the basis of a multiple imputed data set for missing data and adjusted for each individual's pretreatment LVEF, 24-month declines in LVEF averaged 3.3±0.6 percentage points and 3.2±0.7 percentage points, for those randomly assigned to placebo versus statins, respectively (P=0.93). Across both treatment arms, similar percentages of individuals experienced changes of more than 10 percentage points in LVEF, LV strain, LV mass, cognition, and inflammation biomarkers, including among those with greater than 90% drug compliance. CONCLUSIONS: In patients with breast cancer and lymphoma with no existing indication for statin therapy, prospective statin administration did not affect LVEF declines 2 years after doxorubicin. (Funded by the National Institutes of Health; ClinicalTrials.gov number, NCT01988571.).

A planning quality evaluation tool for prostate adaptive IMRT based on machine learning.

PURPOSE: To ensure plan quality for adaptive IMRT of the prostate, we developed a quantitative evaluation tool using a machine learning approach. This tool generates dose volume histograms (DVHs) of organs-at-risk (OARs) based on prior plans as a reference, to be compared with the adaptive plan derived from fluence map deformation. METHODS: Under the same configuration using seven-field 15 MV photon beams, DVHs of OARs (bladder and rectum) were estimated based on anatomical information of the patient and a model learned from a database of high quality prior plans. In this study, the anatomical information was characterized by the organ volumes and distance-to-target histogram (DTH). The database consists of 198 high quality prostate plans and was validated with 14 cases outside the training pool. Principal component analysis (PCA) was applied to DVHs and DTHs to quantify their salient features. Then, support vector regression (SVR) was implemented to establish the correlation between the features of the DVH and the anatomical information. RESULTS: DVH/DTH curves could be characterized sufficiently just using only two or three truncated principal components, thus, patient anatomical information was quantified with reduced numbers of variables. The evaluation of the model using the test data set demonstrated its accuracy approximately 80% in prediction and effectiveness in improving ART planning quality. CONCLUSIONS: An adaptive IMRT plan quality evaluation tool based on machine learning has been developed, which estimates OAR sparing and provides reference in evaluating ART.

A data-driven approach to optimal beam/arc angle selection for liver stereotactic body radiation therapy treatment planning.

BACKGROUND: Stereotactic body radiation therapy (SBRT) for liver cancer has shown promising therapeutic effects. Effective treatment relies not only on the precise delivery provided by image-guided radiation therapy (IGRT) but also high dose gradient formed around the treatment volume to spare functional liver tissue, which is highly dependent on the beam/arc angle selection. In this study, we aim to develop a decision support model to learn human planner's beam navigation approach for beam angle/arc angle selection for liver SBRT. METHODS: A total of 27 liver SBRT/HIGRT patients (10 IMRT, 17 VMAT/DCA) were included in this study. A dosimetric budget index was defined for each beam angle/control point considering dose penetration through the patient body and liver tissue. Optimal beam angle setting (beam angles for IMRT and start/terminal angle for VMAT/DCA) was determined by minimizing the loss function defined as the sum of total dosimetric budget index and beam span penalty function. Leave-one-out validation was exercised on all 27 cases while weighting coefficients in the loss function was tuned in nested cross validation. To compare the efficacy of the model, a model plan was generated using automatically generated beam setting while retaining the original optimization constraints in the clinical plan. Model plan was normalized to the same planning target volume (PTV) V100% as the clinical plans. Dosimetric endpoints including PTV D98%, D2%, liver V20Gy and total MU were compared between two plan groups. Wilcoxon Signed-Rank test was performed with the null hypothesis being that no difference exists between two plan groups. RESULTS: Beam setting prediction was instantaneous. Mean PTV D98% was 91.3% and 91.3% (P=0.566), while mean PTV D2% was 107.9% and 108.1% (P=0.164) for clinical plan and model plan respectively. Liver V20Gy showed no significant difference (P=0.590) with 23.3% for clinical plan and 23.4% for the model plan. Total MU is comparable (P=0.256) between the clinical plan (avg. 2,389.6) and model plan (avg. 2,319.6). CONCLUSIONS: The evidence driven beam setting model yielded similar plan quality as hand-crafted clinical plan. It is capable of capturing human's knowledge in beam selection decision making. This model could facilitate decision making for beam angle selection while eliminating lengthy trial-and-error process of adjusting beam setting during liver SBRT treatment planning.

Insights of an AI agent via analysis of prediction errors: a case study of fluence map prediction for radiation therapy planning.

Purpose.We have previously reported an artificial intelligence (AI) agent that automatically generates intensity-modulated radiation therapy (IMRT) plans via fluence map prediction, by-passing inverse planning. This AI agent achieved clinically comparable quality for prostate cases, but its performance on head-and-neck patients leaves room for improvement. This study aims to collect insights of the deep-learning-based (DL-based) fluence map prediction model by systematically analyzing its prediction errors.Methods.From the modeling perspective, the DL model's output is the fluence maps of IMRT plans. However, from the clinical planning perspective, the plan quality evaluation should be based on the clinical dosimetric criteria such as dose-volume histograms. To account for the complex and non-intuitive relationships between fluence map prediction errors and the corresponding dose distribution changes, we propose a novel error analysis approach that systematically examines plan dosimetric changes that are induced by varying amounts of fluence prediction errors. We investigated four decomposition modes of model prediction errors. The two spatial domain decompositions are based on fluence intensity and fluence gradient. The two frequency domain decompositions are based on Fourier-space banded frequency rings and Fourier-space truncated low-frequency disks. The decomposed error was analyzed for its impact on the resulting plans' dosimetric metrics. The analysis was conducted on 15 test cases spared from the 200 training and 16 validation cases used to train the model.Results.Most planning target volume metrics were significantly correlated with most error decompositions. The Fourier space disk radii had the largest Spearman's coefficients. The low-frequency region within a disk of ∼20% Fourier space contained most of errors that impact overall plan quality.Conclusions.This study demonstrates the feasibility of using fluence map prediction error analysis to understand the AI agent's performance. Such insights will help fine-tune the DL models in architecture design and loss function selection.