Justin J. Boutilier

Optimizing a Drone Network to Deliver Automated External Defibrillators

Background: Public access defibrillation programs can improve survival after out-of-hospital cardiac arrest, but automated external defibrillators (AEDs) are rarely available for bystander use at the scene. Drones are an emerging technology that can deliver an AED to the scene of an out-of-hospital cardiac arrest for bystander use. We hypothesize that a drone network designed with the aid of a mathematical model combining both optimization and queuing can reduce the time to AED arrival. Methods: We applied our model to 53 702 out-of-hospital cardiac arrests that occurred in the 8 regions of the Toronto Regional RescuNET between January 1, 2006, and December 31, 2014. Our primary analysis quantified the drone network size required to deliver an AED 1, 2, or 3 minutes faster than historical median 911 response times for each region independently. A secondary analysis quantified the reduction in drone resources required if RescuNET was treated as a large coordinated region. Results: The region-specific analysis determined that 81 bases and 100 drones would be required to deliver an AED ahead of median 911 response times by 3 minutes. In the most urban region, the 90th percentile of the AED arrival time was reduced by 6 minutes and 43 seconds relative to historical 911 response times in the region. In the most rural region, the 90th percentile was reduced by 10 minutes and 34 seconds. A single coordinated drone network across all regions required 39.5% fewer bases and 30.0% fewer drones to achieve similar AED delivery times. Conclusions: An optimized drone network designed with the aid of a novel mathematical model can substantially reduce the AED delivery time to an out-of-hospital cardiac arrest event.

Models for predicting objective function weights in prostate cancer IMRT

PURPOSE To develop and evaluate the clinical applicability of advanced machine learning models that simultaneously predict multiple optimization objective function weights from patient geometry for intensity-modulated radiation therapy of prostate cancer. METHODS A previously developed inverse optimization method was applied retrospectively to determine optimal objective function weights for 315 treated patients. The authors used an overlap volume ratio (OV) of bladder and rectum for different PTV expansions and overlap volume histogram slopes (OVSR and OVSB for the rectum and bladder, respectively) as explanatory variables that quantify patient geometry. Using the optimal weights as ground truth, the authors trained and applied three prediction models: logistic regression (LR), multinomial logistic regression (MLR), and weighted K-nearest neighbor (KNN). The population average of the optimal objective function weights was also calculated. RESULTS The OV at 0.4 cm and OVSR at 0.1 cm features were found to be the most predictive of the weights. The authors observed comparable performance (i.e., no statistically significant difference) between LR, MLR, and KNN methodologies, with LR appearing to perform the best. All three machine learning models outperformed the population average by a statistically significant amount over a range of clinical metrics including bladder/rectum V53Gy, bladder/rectum V70Gy, and dose to the bladder, rectum, CTV, and PTV. When comparing the weights directly, the LR model predicted bladder and rectum weights that had, on average, a 73% and 74% relative improvement over the population average weights, respectively. The treatment plans resulting from the LR weights had, on average, a rectum V70Gy that was 35% closer to the clinical plan and a bladder V70Gy that was 29% closer, compared to the population average weights. Similar results were observed for all other clinical metrics. CONCLUSIONS The authors demonstrated that the KNN and MLR weight prediction methodologies perform comparably to the LR model and can produce clinical quality treatment plans by simultaneously predicting multiple weights that capture trade-offs associated with sparing multiple OARs.

Knowledge-Based Automated Planning for Oropharyngeal Cancer

Purpose The purpose of this study was to automatically generate radiation therapy plans for oropharynx patients by combining knowledge‐based planning (KBP) predictions with an inverse optimization (IO) pipeline. Methods We developed two KBP approaches, the bagging query (BQ) method and the generalized principal component analysis‐based (gPCA) method, to predict achievable dose–volume histograms (DVHs). These approaches generalize existing methods by predicting physically feasible organ‐at‐risk (OAR) and target DVHs in sites with multiple targets. Using leave‐one‐out cross validation, we applied both models to a large dataset of 217 oropharynx patients. The predicted DVHs were input into an IO pipeline that generated treatment plans (BQ and gPCA plans) via an intermediate step that estimated objective function weights for an inverse planning model. The KBP predictions were compared to the clinical DVHs for benchmarking. To assess the complete pipeline, we compared the BQ and gPCA plans to both the predictions and clinical plans. To isolate the effect of the KBP predictions, we put clinical DVHs through the IO pipeline to produce clinical inverse optimized (CIO) plans. This approach also allowed us to estimate the complexity of the clinical plans. The BQ and gPCA plans were benchmarked against the CIO plans using DVH differences and clinical planning criteria. Iso‐complexity plans (relative to CIO) were also generated and evaluated. Results The BQ method tended to predict that less dose is delivered than what was observed in the clinical plans while the gPCA predictions were more similar to clinical DVHs. Both populations of KBP predictions were reproduced with inverse plans to within a median DVH difference of 3 Gy. Clinical planning criteria for OARs were satisfied most frequently by the BQ plans (74.4%), by 6.3% points more than the clinical plans. Meanwhile, target criteria were satisfied most frequently by the gPCA plans (90.2%), and by 21.2% points more than clinical plans. However, once the complexity of the plans was constrained to that of the CIO plans, the performance of the BQ plans degraded significantly. In contrast, the gPCA plans still satisfied more clinical criteria than both the clinical and CIO plans, with the most notable improvement being in target criteria. Conclusion Our automated pipeline can successfully use DVH predictions to generate high‐quality plans without human intervention. Between the two KBP methods, gPCA plans tend to achieve comparable performance as clinical plans, even when controlling for plan complexity, whereas BQ plans tended to underperform.

Applications of machine learning algorithms to predict therapeutic outcomes in depression: A meta-analysis and systematic review

BACKGROUND No previous study has comprehensively reviewed the application of machine learning algorithms in mood disorders populations. Herein, we qualitatively and quantitatively evaluate previous studies of machine learning-devised models that predict therapeutic outcomes in mood disorders populations. METHODS We searched Ovid MEDLINE/PubMed from inception to February 8, 2018 for relevant studies that included adults with bipolar or unipolar depression; assessed therapeutic outcomes with a pharmacological, neuromodulatory, or manual-based psychotherapeutic intervention for depression; applied a machine learning algorithm; and reported predictors of therapeutic response. A random-effects meta-analysis of proportions and meta-regression analyses were conducted. RESULTS We identified 639 records: 75 full-text publications were assessed for eligibility; 26 studies (n=17,499) and 20 studies (n=6325) were included in qualitative and quantitative review, respectively. Classification algorithms were able to predict therapeutic outcomes with an overall accuracy of 0.82 (95% confidence interval [CI] of [0.77, 0.87]). Pooled estimates of classification accuracy were significantly greater (p < 0.01) in models informed by multiple data types (e.g., composite of phenomenological patient features and neuroimaging or peripheral gene expression data; pooled proportion [95% CI] = 0.93[0.86, 0.97]) when compared to models with lower-dimension data types (pooledproportion=0.68[0.62,0.74]to0.85[0.81,0.88]). LIMITATIONS Most studies were retrospective; differences in machine learning algorithms and their implementation (e.g., cross-validation, hyperparameter tuning); cannot infer importance of individual variables fed into learning algorithm. CONCLUSIONS Machine learning algorithms provide a powerful conceptual and analytic framework capable of integrating multiple data types and sources. An integrative approach may more effectively model neurobiological components as functional modules of pathophysiology embedded within the complex, social dynamics that influence the phenomenology of mental disorders.

WE-AB-209-11: Prostate Cancer Treatment Planning: Sensitivity and Representative Objective Function Weights

PURPOSE To develop an automated planning methodology that exploits patient sensitivity to objective function weights. METHODS Given a treatment plan, we first create an acceptable treatment region that encompasses a set of treatment plans with similar clinical performance (e.g., +/-1% at V70Gy). We use inverse optimization to map this region in criterion space to the weight space and find a corresponding region of acceptable weight vectors (W). The shape and size of W describes how sensitive a patient is to perturbations in objective function weights. To exploit the information encoded by these regions, we approximate W for each patient by a polyhedron and we cluster patients using a novel integer programming model with cluster sizes from k=1,2,…,10. Each cluster centroid is a representative objective function weight vector and we use these weight vectors to generate k treatment plans for each patient (AUTO plans). Using 315 prostate cancer plans, we determine the number of patients that would have received an improved treatment plan using our automated approach. RESULTS Clustering patients into five groups produced a global set of representative weights such that for 88% of patients there exists at least one AUTO plan that improves upon the clinical treatment plan in terms of organ-at-risk mean dose and clinical acceptability criteria satisfaction (i.e., V54Gy50% and V70Gy30%). The AUTO plans provided bladder or rectum mean dose improvement over the clinical treatment plans for 296 (94%) patients, bladder mean dose improvement for 185 (59%) patients, rectum mean dose improvement for 273 (87%) patients, and mean dose improvements for both bladder and rectum in 162 (51%) patients. The AUTO plans provided fewer violations for bladder/rectum V54Gy50% and slightly more for bladder/rectum V70Gy30% when compared to clinical plans. CONCLUSION A method combining inverse optimization and clustering automatically produces prostate treatment plans for 88% of patients.

MO-G-201-02: Comparing Sample Size Requirements for Knowledge-Based Treatment Planning

PURPOSE To compare how training set size affects the accuracy of a knowledge-based planning (KBP) model applied to prostate and head and neck (HN) cancer. METHODS We selected a KBP model from the literature that uses distance-to-target histograms and organ volumes to predict an achievable dose-volume-histogram (DVH) curve for each organ-at-risk (OAR). We trained both the prostate and HN model using training set sizes of n=10, 20, 30, 50,75, and 100. We set aside 100 randomly selected treatment plans from each of the two respective cohorts of 218 to serve as a validation set for all experiments. For each value of n, we randomly selected 100 different training sets with replacement from the remaining 118 plans. Each of the 100 training sets was used to train a model for each value of n and for both prostate and HN. To evaluate the models we predicted DVH curves for each of the 100 plans in the validation set. To estimate the minimum required sample size, we used the rank-sum test to determine if the median error for each sample size from 10 to 75 was equal to the median error for the maximum sample size of 100. RESULTS In general, larger sample sizes were required for HN compared to prostate. For prostate, a minimum training set size of 30 plans was needed to accurately predict the bladder DVH, while at least 75 plans were needed for the rectum. For HN, the minimum training set size was 100 for the larynx esophagus and spinal cord, 75 for the left parotid and mandible, and only 50 for the right parotid. CONCLUSION The minimum sample size required for accurate treatment plan generation using KBP is OAR and site dependent. Adequate sample sizes are essential for successful clinical implementation of KBP models. This research was funded in part by the Natural Sciences and Engineering Research Council of Canada.

SU-F-BRD-01: A Logistic Regression Model to Predict Objective Function Weights in Prostate Cancer IMRT

PURPOSE To develop a statistical model that predicts optimization objective function weights from patient geometry for intensity-modulation radiotherapy (IMRT) of prostate cancer. METHODS A previously developed inverse optimization method (IOM) is applied retrospectively to determine optimal weights for 51 treated patients. We use an overlap volume ratio (OVR) of bladder and rectum for different PTV expansions in order to quantify patient geometry in explanatory variables. Using the optimal weights as ground truth, we develop and train a logistic regression (LR) model to predict the rectum weight and thus the bladder weight. Post hoc, we fix the weights of the left femoral head, right femoral head, and an artificial structure that encourages conformity to the population average while normalizing the bladder and rectum weights accordingly. The population average of objective function weights is used for comparison. RESULTS The OVR at 0.7cm was found to be the most predictive of the rectum weights. The LR model performance is statistically significant when compared to the population average over a range of clinical metrics including bladder/rectum V53Gy, bladder/rectum V70Gy, and mean voxel dose to the bladder, rectum, CTV, and PTV. On average, the LR model predicted bladder and rectum weights that are both 63% closer to the optimal weights compared to the population average. The treatment plans resulting from the LR weights have, on average, a rectum V70Gy that is 35% closer to the clinical plan and a bladder V70Gy that is 43% closer. Similar results are seen for bladder V54Gy and rectum V54Gy. CONCLUSION Statistical modelling from patient anatomy can be used to determine objective function weights in IMRT for prostate cancer. Our method allows the treatment planners to begin the personalization process from an informed starting point, which may lead to more consistent clinical plans and reduce overall planning time.