Assessment of Tissue Toxicity Risk in Breast Radiotherapy using Bayesian Networks

PURPOSE

The purpose of this analysis is to predict worsening post-treatment normal tissue toxicity in patients undergoing accelerated partial breast irradiation (APBI) therapy and to quantitatively identify which diagnostic, anatomical, and dosimetric features are contributing to these outcomes.

METHODS

A retrospective study of APBI treatments was performed using 32 features pertaining to various stages of the patient's treatment journey. These features were used to inform and construct a Bayesian network (BN) based on both statistical analysis of feature distributions and relative clinical importance. The target feature for prediction was defined as a measurable worsening of telangiectasia, subcutaneous tissue induration, or fibrosis when compared against the observed baseline. Parameter learning for the network was performed using data from the 299 patients included in the ACCEL trial and predictive performance was measured. Feature importance for the BN was quantified using a novel information-theoretic approach.

RESULTS

Cross validated performance of the BN for predicting toxicity was consistently higher when compared against conventional machine learning (ML) techniques. The measured BN receiver operating characteristic area under the curve was 0.960±0.013 against the best ML result of 0.942±0.021 using 5-fold cross validation with separate test data across 100 trials. The volume of the clinical target volume, gross target volume, and baseline toxicity measurements were found to have the highest feature importance and mutual dependence with normal tissue toxicity in the network, representing the strongest contribution to patient outcomes.

CONCLUSIONS

The BN outperformed conventional ML techniques in predicting tissue toxicity outcomes and provided deeper insight into which features are contributing to these outcomes. This article is protected by copyright. All rights reserved.

Can We Estimate Short- and Intermediate-term Survival in Patients Undergoing Surgery for Metastatic Bone Disease?

BACKGROUND

Objective means of estimating survival can be used to guide surgical decision-making and to risk-stratify patients for clinical trials. Although a free, online tool ( www.pathfx.org ) can estimate 3- and 12-month survival, recent work, including a survey of the Musculoskeletal Tumor Society, indicated that estimates at 1 and 6 months after surgery also would be helpful. Longer estimates help justify the need for more durable and expensive reconstructive options, and very short estimates could help identify those who will not survive 1 month and should not undergo surgery. Thereby, an important use of this tool would be to help avoid unsuccessful and expensive surgery during the last month of life.

QUESTIONS/PURPOSES

We seek to provide a reliable, objective means of estimating survival in patients with metastatic bone disease. After generating models to derive 1- and 6-month survival estimates, we determined suitability for clinical use by applying receiver operator characteristic (ROC) (area under the curve [AUC] > 0.7) and decision curve analysis (DCA), which determines whether using PATHFx can improve outcomes, but also discerns in which kinds of patients PATHFx should not be used.

METHODS

We used two, existing, skeletal metastasis registries chosen for their quality and availability. Data from Memorial Sloan-Kettering Cancer Center (training set, n = 189) was used to develop two Bayesian Belief Networks trained to estimate the likelihood of survival at 1 and 6 months after surgery. Next, data from eight major referral centers across Scandinavia (n = 815) served as the external validation set-that is, as a means to test model performance in a different patient population. The diversity of the data between the training set from Memorial Sloan-Kettering Cancer Center and the Scandinavian external validation set is important to help ensure the models are applicable to patients in various settings with differing demographics and treatment philosophies. We considered disease-specific, laboratory, and demographic information, and the surgeon's estimate of survival. For each model, we calculated the area under the ROC curve (AUC) as a metric of discriminatory ability and the Net Benefit using DCA to determine whether the models were suitable for clinical use.

RESULTS

On external validation, the AUC for the 1- and 6-month models were 0.76 (95% CI, 0.72-0.80) and 0.76 (95% CI, 0.73-0.79), respectively. The models conferred a positive net benefit on DCA, indicating each could be used rather than assume all patients or no patients would survive greater than 1 or 6 months, respectively.

CONCLUSIONS

Decision analysis confirms that the 1- and 6-month Bayesian models are suitable for clinical use.

CLINICAL RELEVANCE

These data support upgrading www.pathfx.org with the algorithms described above, which is designed to guide surgical decision-making, and function as a risk stratification method in support of clinical trials. This updating has been done, so now surgeons may use any web browser to generate survival estimates at 1, 3, 6, and 12 months after surgery, at no cost. Just as short estimates of survival help justify palliative therapy or less-invasive approaches to stabilization, more favorable survival estimates at 6 or 12 months are used to justify more durable, complicated, and expensive reconstructive options.

Assessing Chemical Mixtures and Human Health: Use of Bayesian Belief Net Analysis

BACKGROUND: Despite humans being exposed to complex chemical mixtures, much of the available research continues to focus on a single compound or metabolite or a select subgroup of compounds inconsistent with the nature of human exposure. Uncertainty regarding how best to model chemical mixtures coupled with few analytic approaches remains a formidable challenge and served as the impetus for study. OBJECTIVES: To identify the polychlorinated biphenyl (PCB) congener(s) within a chemical mixture that was most associated with an endometriosis diagnosis using novel graphical modeling techniques. METHODS: Bayesian Belief Network (BBN) models were developed and empirically assessed in a cohort comprising 84 women aged 18-40 years who underwent a laparoscopy or laparotomy between 1999 and 2000; 79 (94%) women had serum concentrations for 68 PCB congeners quantified. Adjusted odds ratios (AOR) for endometriosis were estimated for individual PCB congeners using BBN models. RESULTS: PCB congeners #114 (AOR = 3.01; 95% CI = 2.25, 3.77) and #136 (AOR = 1.79; 95% CI = 1.03, 2.55) were associated with an endometriosis diagnosis. Combinations of mixtures inclusive of PCB #114 were all associated with higher odds of endometriosis, underscoring its potential relation with endometriosis. CONCLUSIONS: BBN models identified PCB congener 114 as the most influential congener for the odds of an endometriosis diagnosis in the context of a 68 congener chemical mixture. BBN models offer investigators the opportunity to assess which compounds within a mixture may drive a human health effect.