Visual Assistance in Development and Validation of Bayesian Networks for Clinical Decision Support.

The development and validation of Clinical Decision Support Models (CDSM) based on Bayesian networks (BN) is commonly done in a collaborative work between medical researchers providing the domain expertise and computer scientists developing the decision support model. Although modern tools provide facilities for data-driven model generation, domain experts are required to validate the accuracy of the learned model and to provide expert knowledge for fine-tuning it while computer scientists are needed to integrate this knowledge in the learned model (hybrid modeling approach). This generally time-expensive procedure hampers CDSM generation and updating. To address this problem, we developed a novel interactive visual approach allowing medical researchers with less knowledge in CDSM to develop and validate BNs based on domain specific data mainly independently and thus, diminishing the need for an additional computer scientist. In this context, we abstracted and simplified the common workflow in BN development as well as adjusted the workflow to medical experts' needs. We demonstrate our visual approach with data of endometrial cancer patients and evaluated it with six medical researchers who are domain experts in the gynecological field.

Improved preoperative risk stratification in endometrial carcinoma patients: external validation of the ENDORISK Bayesian network model in a large population-based case series.

PURPOSE: Preoperative risk stratification of newly diagnosed endometrial carcinoma (EC) patients has been hindered by only moderate prediction performance for many years. Recently ENDORISK, a Bayesian network model, showed high predictive performance. It was the aim of this study to validate ENDORISK by applying the model to a population-based case series of EC patients. METHODS: ENDORISK was applied to a retrospective cohort of women surgically treated for EC from 2003 to 2013. Prediction accuracy for LNM as well as 5-year DSS was investigated. The model's overall performance was quantified by the Brier score, discriminative performance by area under the curve (AUC). RESULTS: A complete dataset was evaluable from 247 patients. 78.1% cases were endometrioid histotype. The majority of patients (n = 156;63.2%) had stage IA disease. Overall, positive lymph nodes were found in 20 (8.1%) patients. Using ENDORISK predicted probabilities, most (n = 156;63.2%) patients have been assigned to low or very low risk group with a false-negative rate of 0.6%. AUC for LNM prediction was 0.851 [95% confidence interval (CI) 0.761-0.941] with a Brier score of 0.06. For 5-year DSS the AUC was 0.698 (95% CI 0.595-0.800) as Brier score has been calculated 0.09. CONCLUSIONS: We were able to successfully validate ENDORISK for prediction of LNM and 5-year DSS. Next steps will now have to focus on ENDORISK performance in daily clinical practice. In addition, incorporating TCGA-derived molecular subtypes will be of key importance for future extended use. This study may support further promoting of data-based decision-making tools for personalized treatment of EC.

External validation study of endometrial cancer preoperative risk stratification model (ENDORISK).

Introduction: Among industrialized countries, endometrial cancer is a common malignancy with generally an excellent outcome. To personalize medicine, we ideally compile as much information as possible concerning patient prognosis prior to effecting an appropriate treatment decision. Endometrial cancer preoperative risk stratification (ENDORISK) is a machine learning-based computational Bayesian networks model that predicts lymph node metastasis and 5-year disease-specific survival potential with percentual probability. Our objective included validating ENDORISK effectiveness in our patient cohort, assessing its application in the current use of sentinel node biopsy, and verifying its accuracy in advanced stages. Methods: The ENDORISK model was evaluated with a retrospective cohort of 425 patients from the University Hospital Brno, Czech Republic. Two hundred ninety-nine patients were involved in our disease-specific survival analysis; 226 cases with known lymph node status were available for lymph node metastasis analysis. Patients were included undergoing either pelvic lymph node dissection (N = 84) or sentinel node biopsy (N =70) to explore the accuracy of both staging procedures. Results: The area under the curve was 0.84 (95% confidence interval [CI], 0.77-0.9) for lymph node metastasis analysis and 0.86 (95% CI, 0.79-0.93) for 5-year disease-specific survival evaluation, indicating quite positive concordance between prediction and reality. Calibration plots to visualize results demonstrated an outstanding predictive value for low-risk cancers (grades 1-2), whereas outcomes were underestimated among high-risk patients (grade 3), especially in disease-specific survival. This phenomenon was even more obvious when patients were subclassified according to FIGO clinical stages. Conclusions: Our data confirmed ENDORISK model's laudable predictive ability, particularly among patients with a low risk of lymph node metastasis and expected favorable survival. For high-risk and/or advanced stages, the ENDORISK network needs to be additionally trained/improved.

Preoperative risk stratification in endometrial cancer (ENDORISK) by a Bayesian network model: A development and validation study.

BACKGROUND: Bayesian networks (BNs) are machine-learning-based computational models that visualize causal relationships and provide insight into the processes underlying disease progression, closely resembling clinical decision-making. Preoperative identification of patients at risk for lymph node metastasis (LNM) is challenging in endometrial cancer, and although several biomarkers are related to LNM, none of them are incorporated in clinical practice. The aim of this study was to develop and externally validate a preoperative BN to predict LNM and outcome in endometrial cancer patients. METHODS AND FINDINGS: Within the European Network for Individualized Treatment of Endometrial Cancer (ENITEC), we performed a retrospective multicenter cohort study including 763 patients, median age 65 years (interquartile range [IQR] 58-71), surgically treated for endometrial cancer between February 1995 and August 2013 at one of the 10 participating European hospitals. A BN was developed using score-based machine learning in addition to expert knowledge. Our main outcome measures were LNM and 5-year disease-specific survival (DSS). Preoperative clinical, histopathological, and molecular biomarkers were included in the network. External validation was performed using 2 prospective study cohorts: the Molecular Markers in Treatment in Endometrial Cancer (MoMaTEC) study cohort, including 446 Norwegian patients, median age 64 years (IQR 59-74), treated between May 2001 and 2010; and the PIpelle Prospective ENDOmetrial carcinoma (PIPENDO) study cohort, including 384 Dutch patients, median age 66 years (IQR 60-73), treated between September 2011 and December 2013. A BN called ENDORISK (preoperative risk stratification in endometrial cancer) was developed including the following predictors: preoperative tumor grade; immunohistochemical expression of estrogen receptor (ER), progesterone receptor (PR), p53, and L1 cell adhesion molecule (L1CAM); cancer antigen 125 serum level; thrombocyte count; imaging results on lymphadenopathy; and cervical cytology. In the MoMaTEC cohort, the area under the curve (AUC) was 0.82 (95% confidence interval [CI] 0.76-0.88) for LNM and 0.82 (95% CI 0.77-0.87) for 5-year DSS. In the PIPENDO cohort, the AUC for 5-year DSS was 0.84 (95% CI 0.78-0.90). The network was well-calibrated. In the MoMaTEC cohort, 249 patients (55.8%) were classified with <5% risk of LNM, with a false-negative rate of 1.6%. A limitation of the study is the use of imputation to correct for missing predictor variables in the development cohort and the retrospective study design. CONCLUSIONS: In this study, we illustrated how BNs can be used for individualizing clinical decision-making in oncology by incorporating easily accessible and multimodal biomarkers. The network shows the complex interactions underlying the carcinogenetic process of endometrial cancer by its graphical representation. A prospective feasibility study will be needed prior to implementation in the clinic.

On the interplay of machine learning and background knowledge in image interpretation by Bayesian networks.

OBJECTIVES: To obtain a balanced view on the role and place of expert knowledge and learning methods in building Bayesian networks for medical image interpretation. METHODS AND MATERIALS: The interpretation of mammograms was selected as the example medical image interpretation problem. Medical image interpretation has its own common standards and procedures. The impact of these on two complementary methods for Bayesian network construction was explored. Firstly, methods for the discretisation of continuous features were investigated, yielding multinomial distributions that were compared to the original Gaussian probabilistic parameters of the network. Secondly, the structure of a manually constructed Bayesian network was tested by structure learning from image data. The image data used for the research came from screening mammographic examinations of 795 patients, of whom 344 were cancerous. RESULTS: The experimental results show that there is an interesting interplay of machine learning results and background knowledge in medical image interpretation. Networks with discretised data lead to better classification performance (increase in the detected cancers of up to 11.7%), easier interpretation, and a better fit to the data in comparison to the expert-based Bayesian network with Gaussian probabilistic parameters. Gaussian probability distributions are often used in medical image interpretation because of the continuous nature of many of the image features. The structures learnt supported many of the expert-originated relationships but also revealed some novel relationships between the mammographic features. Using discretised features and performing structure learning on the mammographic data has further improved the cancer detection performance of up to 17% compared to the manually constructed Bayesian network model. CONCLUSION: Finding the right balance between expert knowledge and data-derived knowledge, both at the level of network structure and parameters, is key to using Bayesian networks for medical image interpretation. A balanced approach to building Bayesian networks for image interpretation yields more accurate and understandable Bayesian network models.

A probabilistic framework for image information fusion with an application to mammographic analysis.

The recent increased interest in information fusion methods for solving complex problem, such as in image analysis, is motivated by the wish to better exploit the multitude of information, available from different sources, to enhance decision-making. In this paper, we propose a novel method, that advances the state of the art of fusing image information from different views, based on a special class of probabilistic graphical models, called causal independence models. The strength of this method is its ability to systematically and naturally capture uncertain domain knowledge, while performing information fusion in a computationally efficient way. We examine the value of the method for mammographic analysis and demonstrate its advantages in terms of explicit knowledge representation and accuracy (increase of at least 6.3% and 5.2% of true positive detection rates at 5% and 10% false positive rates) in comparison with previous single-view and multi-view systems, and benchmark fusion methods such as naïve Bayes and logistic regression.

Using local context information to improve automatic mammographic mass detection.

Despite their promising application, current Computer-Aided Detection (CAD) systems face difficulties, especially in the detection of malignant masses -a major mammographic sign for breast cancer. One of the main problems is the large number of false positives prompted, which is a critical issue in screening programs where the number of normal cases is considerably large. A crucial determinant for this problem is the dependence of the CAD output on the single pixel-based locations initially detected. To refine the initial detection step, in this paper, we propose a novel approach by considering the context information between the neighbouring pixel features and classes for every initially detected suspicious location. Our modelling scheme is based on the Conditional Random Field technique and the mammographic features extracted by image processing techniques. In experimental study, we demonstrated the practical application of the approach and we compared its performance to that of a previously developed CAD system. The results demonstrated the superiority of the context modelling in terms of significantly improved accuracy without increase in computation efforts.

Improved mammographic CAD performance using multi-view information: a Bayesian network framework.

Mammographic reading by radiologists requires the comparison of at least two breast projections (views) for the detection and the diagnosis of breast abnormalities. Despite their reported potential to support radiologists, most mammographic computer-aided detection (CAD) systems have a major limitation: as opposed to the radiologist's practice, computerized systems analyze each view independently. To tackle this problem, in this paper, we propose a Bayesian network framework for multi-view mammographic analysis, with main focus on breast cancer detection at a patient level. We use causal-independence models and context modeling over the whole breast represented as links between the regions detected by a single-view CAD system in the two breast projections. The proposed approach is implemented and tested with screening mammograms for 1063 cases of whom 385 had breast cancer. The single-view CAD system is used as a benchmark method for comparison. The results show that our multi-view modeling leads to significantly better performance in discriminating between normal and cancerous patients. We also demonstrate the potential of our multi-view system for selecting the most suspicious cases.

Using model checking for critiquing based on clinical guidelines.

OBJECTIVE: Medical critiquing systems compare clinical actions performed by a physician with a predefined set of actions. In order to provide useful feedback, an important task is to find differences between the actual actions and a set of 'ideal' actions as described by a clinical guideline. In case differences exist, the critiquing system provides insight into the extent to which they are compatible. METHODS AND MATERIAL: We propose a computational method for such critiquing, where the ideal actions are given by a formal model of a clinical guideline, and where the actual actions are derived from real world patient data. We employ model checking to investigate whether a part of the actual treatment is consistent with the guideline. RESULTS: We show how critiquing can be cast in terms of temporal logic, and what can be achieved by using model checking. Furthermore, a method is introduced for off-line computing relevant information which can be exploited during critiquing. The method has been applied to a clinical guideline of breast cancer in conjunction with breast cancer patient data.

Dynamic Bayesian networks as prognostic models for clinical patient management.

Prognostic models in medicine are usually been built using simple decision rules, proportional hazards models, or Markov models. Dynamic Bayesian networks (DBNs) offer an approach that allows for the incorporation of the causal and temporal nature of medical domain knowledge as elicited from domain experts, thereby allowing for detailed prognostic predictions. The aim of this paper is to describe the considerations that must be taken into account when constructing a DBN for complex medical domains and to demonstrate their usefulness in practice. To this end, we focus on the construction of a DBN for prognosis of carcinoid patients, compare performance with that of a proportional hazards model, and describe predictions for three individual patients. We show that the DBN can make detailed predictions, about not only patient survival, but also other variables of interest, such as disease progression, the effect of treatment, and the development of complications. Strengths and limitations of our approach are discussed and compared with those offered by traditional methods.

Selecting treatment strategies with dynamic limited-memory influence diagrams.

OBJECTIVE: The development of dynamic limited-memory influence diagrams as a framework for representing factorized infinite-horizon partially observable Markov decision processes (POMDPs), the introduction of algorithms for their (approximate) solution, and the application to a dynamic decision problem in clinical oncology. MATERIALS AND METHODS: A dynamic limited-memory influence diagram for high-grade carcinoid tumor pathophysiology was developed in collaboration with an expert physician. Three algorithms, known as single policy updating, single rule updating, and simulated annealing have been examined for approximating the optimal treatment strategy from a space of 10(19) possible strategies. RESULTS: Single policy updating proved intractable for finding a treatment strategy for carcinoid tumors. Single rule updating and simulated annealing both found the treatment strategy that is applied by physicians in practice. CONCLUSIONS: Dynamic limited-memory influence diagrams are a suitable framework for the representation of factorized infinite-horizon POMDPs, and the developed algorithms find acceptable solutions under the assumption of limited memory about past observations. The framework allows for finding reasonable treatment strategies for complex dynamic decision problems in medicine.

Predicting carcinoid heart disease with the noisy-threshold classifier.

OBJECTIVE: To predict the development of carcinoid heart disease (CHD), which is a life-threatening complication of certain neuroendocrine tumors. To this end, a novel type of Bayesian classifier, known as the noisy-threshold classifier, is applied. MATERIALS AND METHODS: Fifty-four cases of patients that suffered from a low-grade midgut carcinoid tumor, of which 22 patients developed CHD, were obtained from the Netherlands Cancer Institute (NKI). Eleven attributes that are known at admission have been used to classify whether the patient develops CHD. Classification accuracy and area under the receiver operating characteristics (ROC) curve of the noisy-threshold classifier are compared with those of the naive-Bayes classifier, logistic regression, the decision-tree learning algorithm C4.5, and a decision rule, as formulated by an expert physician. RESULTS: The noisy-threshold classifier showed the best classification accuracy of 72% correctly classified cases, although differences were significant only for logistic regression and C4.5. An area under the ROC curve of 0.66 was attained for the noisy-threshold classifier, and equaled that of the physician's decision-rule. CONCLUSIONS: The noisy-threshold classifier performed favorably to other state-of-the-art classification algorithms, and equally well as a decision-rule that was formulated by the physician. Furthermore, the semantics of the noisy-threshold classifier make it a useful machine learning technique in domains where multiple causes influence a common effect.