A combined neural network and decision trees model for prognosis of breast cancer relapse

doi:10.1016/S0933-3657(02)00086-6

Artificial Intelligence in Medicine

Volume 27, Issue 1, January 2003, Pages 45-63

https://doi.org/10.1016/S0933-3657(02)00086-6 Get rights and content

Abstract

The prediction of clinical outcome of patients after breast cancer surgery plays an important role in medical tasks such as diagnosis and treatment planning. Different prognostic factors for breast cancer outcome appear to be significant predictors for overall survival, but probably form part of a bigger picture comprising many factors. Survival estimations are currently performed by clinicians using the statistical techniques of survival analysis. In this sense, artificial neural networks are shown to be a powerful tool for analysing datasets where there are complicated non-linear interactions between the input data and the information to be predicted. This paper presents a decision support tool for the prognosis of breast cancer relapse that combines a novel algorithm TDIDT (control of induction by sample division method, CIDIM), to select the most relevant prognostic factors for the accurate prognosis of breast cancer, with a system composed of different neural networks topologies that takes as input the selected variables in order for it to reach good correct classification probability. In addition, a new method for the estimate of Bayes’ optimal error using the neural network paradigm is proposed. Clinical–pathological data were obtained from the Medical Oncology Service of the Hospital Clı́nico Universitario of Málaga, Spain. The results show that the proposed system is an useful tool to be used by clinicians to search through large datasets seeking subtle patterns in prognostic factors, and that may further assist the selection of appropriate adjuvant treatments for the individual patient.

Introduction

Prediction tasks are among the most interesting activities in which to implement intelligent systems. Specifically, prediction is an attempt to accurately forecast the outcome of a specific situation, using as input information obtained from a concrete set of variables that potentially describe the situation.

A problem often faced in clinical medicine is how to reach a conclusion about the prognosis of cancer patients when presented with complex clinical and prognostic information, since specialists usually make decisions based on a simple dichotomization of variables into a favourable and unfavourable classification [18]. As we enter the new millennium, treatment modalities exist for many solid tumour types and their use is well established. Nevertheless, offset against this is the toxicity of some treatments. As there is a real risk of mortality associated with treatment, it is vital to have the possibility of offering different therapies depending on the patients. In this sense, the likelihood that the patient will suffer a recurrence of her disease is very important, so that the risks and expected benefits of specific therapies can be compared.

This work analyses, on the one hand, the decision-making process existing when patients with primary breast cancer should receive a certain therapy to remove the primary tumour. On the other hand, different prognostic factors appear to be significant predictors for overall survival, but probably form part of a bigger picture comprising many, inter-related factors [11]. In order to investigate this hypothesis, studies looking at a large number of potential prognostic factors are needed. To further complicate matters, these relationships may well be non-linear in nature. These form the major difficulties in such studies. Furthermore, the statistical analysis of large datasets using standard methodologies is cumbersome and limited, especially in the case of non-linear relationships.

Among prognostic modelling techniques that induce models from medical data, survival analysis methods are specific both in terms of modelling and the type of data required. Survival models attempt to determine the probability of the event occurring within a specific time, which requires classification models that classify either the occurrence or non-occurrence of the event and optionally model the outcome probabilities. Several tools successfully used in the construction of medical prognosis models have been proposed by the machine learning community [17], [34].

Neural networks are a form of artificial intelligence that have found application in a wide range of problems [10], [20], [24] and have given, in many cases, superior results to standard statistical models [33]. Baxt [4] demonstrated the predictive reliability of an artificial neural networks model in medical diagnosis. In this case, we utilise the ability of neural networks to recognise complex and highly non-linear relationships, such as are likely to characterise medical circumstances.

Some authors [14], [30] have modelled systems for outcome prediction in post-surgery breast and lung carcinoma patients using neural networks to perform survival analysis. This type of modelling manages the problem of censored data handling that arises when the event related to the censor variable—normally included in the survival data (like death or recurrence of a disease)—has not occurred during the follow-up period for a patient, although the event may eventually occur. These authors have solved the problem by using different survival estimators to handle censored data for patients. This would imply that prognostic factors—for example, in breast cancer with adjuvant therapy after surgery—are not time-dependent, but this is not really true. That is, the strength of the prognostic factor is not the same for different time intervals. Different techniques for survival estimation, such as Kaplan–Meier analysis [15] and Cox Regression modelling [6] assume that the strength of a prognostic factor does not change over time. In addition, the existence of a “peak” of recurrence in the distribution of relapse probability [2] demonstrates that the recurrence probability is not the same over time. In this sense, if these statistical techniques are not appropriate to solve this problem, a possible solution would be to incorporate the whole set of prognostic factors pre-selected by medical experts (Section 3.1) as input to the neural networks system. This would involve removing all the patients with censor data; however, the cardinality of the resulting patient data vectors set would then become too small to constitute a significant representation of this problem.

This work proposes a new system approach based on: (1) specific topologies of neural networks for different time intervals during the follow-up time of the patients, considering the events occurring in different intervals as different problems; and (2) decision trees, useful in understanding the underlying relationships in breast cancer data, for selecting the most important prognostic factors corresponding to every time interval. This is not the first attempt to combine decision trees and neural networks [1], [7], but it does present different ways of integrating them.

In addition, we introduce a new decision trees algorithm, control of induction by sample division method (CIDIM), for reducing the number of rules and improving the selection of attributes from the database to become significant prognostic factors. Furthermore, a new upper-bound estimate of the problem-difficulty level, based on the correct classification Bayes’ probability, is also proposed.

Section snippets

Breast cancer overview

Breast cancer is a malignant tumour that has developed from cells of the breast. Although scientists know some of the risk factors (i.e. ageing, genetic risk factors, family history, menstrual periods, not having children, obesity) that increase a woman’s chance of developing breast cancer, they do not yet know what causes most breast cancers or exactly how some of these risk factors cause cells to become cancerous. Research is under way to learn more and scientists are making great progress in

Patient data

Data from 1035 patients with breast cancer disease from the Medical Oncology Service of the Hospital Clı́nico Universitario of Málaga, Spain were collected and recorded during the period 1990–2000. Data corresponding to every patient were structured in 85 fields containing information about post-surgical measurements, personal data, and type of treatment. Part of this information regarding patients is not relevant for predicting outcome, so that only 14 independent input variables—pre-selected

Results and discussion

Table 6 shows the number of patients and the selection of prognostic factors corresponding to every time interval (in months) of patients’ follow-up that were selected for training the neural networks system. After processing the patient database through the decision trees system (CIDIM algorithm), certain attributes appear to be the most significant prognostic factors (second column in Table 6) becoming the input to the artificial neural networks system. The decision trees system makes the

Conclusions

This paper presents a decision-support tool for the prognosis of breast cancer relapse using clinical–pathological data. We propose a model that combines a novel algorithm TDIDT (CIDIM), with a system composed of different neural network topologies to approximate Bayes’ optimal error for the prediction of patient relapse after breast cancer surgery. The CIDIM algorithm selects the most relevant prognostic factors for the accurate prognosis of breast cancer, while the neural networks system

Acknowledgements

We would like to thank the referees for their valuable comments and suggestions, and also the Oncology Service staff of the Hospital Clı́nico Universitario of Málaga for their comments and collaboration in this work. This work has been partially supported by the FRESCO project, number PB98-0937-C04-01, of CICYT Spain.

References (34)

W.G. Baxt
Application of neural networks to clinical medicine
Lancet
(1995)
K. Funahashi
Multilayer neural networks and Bayes decision theory
Neural Networks
(1998)
R.P. Gorman et al.
Analysis of hidden units in a layered network trained to classify sonar targets
Neural Networks
(1988)
S. Grumett et al.
Artificial neural networks: a new model for assessing prognostic factors
Ann. Oncol.
(2000)
M.W. Kattan et al.
Experiments to determine whether recursive partitioning (cart) or an artificial neural network overcomes theoretical limitations of Cox proportional hazards regression
Comput Biomed Res
(1998)
P.J.F. Lucas et al.
Prognostic methods in medicine
Artif Intell Med
(1999)
E. Pesonen et al.
Comparison of different neural networks algorithms in the diagnosis of acute apendicitis
Int J Biomed Comput
(1996)
E. Pesonen et al.
Treatment of missing data values in a neural network based decision support system for acute abdominal pain
Artif Intell Med
(1998)
N. Qian et al.
Predicting the secondary structure of globular proteins using neural network models
J Mol Biol
(1988)
B. Zupan et al.
Machine learning for survival analysis: a case study on recurrence of prostate cancer
Artif Intell Med
(2000)

H.A. Abbass et al.

C-Net: a method for generating non-deterministic and dynamic multivariate decision trees

Know. Inform. Syst.

(2001)

Alba E et. al. Estructura del patron de recurrencia en el cancer de mama operable (CMO) tras el tratamiento primario....

S. Amari et al.

Statistical theory of overtraining—is cross-validation asymptotically effective?

Adv. Neural Inform. Process. Syst.

(1996)

W. Buntine et al.

A further comparison of splitting rules for decision-tree induction

Mach. Learn.

(1992)

D.R. Cox

Regression models and life tables

J. R. Stat. Soc.

(1972)

F. D’alche-Buc et al.

Trio learning: a new strategy for building hybrid neural trees

Neural Syst.

(1994)

Duda RO, Hart PE. Pattern classification and scene analysis. New York: Wiley;...

Cited by (182)

Application of the convolutional neural networks and supervised deep-learning methods for osteosarcoma bone cancer detection
2023, Healthcare Analytics
Osteosarcoma is a cancerous tumor that occurs in bones. Although it can occur in any bone, it often occurs in long bones such as arms and legs. The exact cause of this cancerous tumor is still unknown, but according to experts, it occurs due to the deoxyribonucleic acid (DNA) mutations inside the bones. This creates immature, irregular, diseased bone and can destroy healthy body tissue. About 75 out of 100 people who have osteosarcoma can be cured if the cancer is not dispersed to the additional body parts. The bone X-ray is the initial test when a bone tumor is suspected. X-ray and imaging tests are the best way to identify osteosarcoma from the bones. A biopsy is the suggested method that can make a definitive diagnosis. This is a time-consuming and difficult procedure that can be automated. We propose several supervised deep-learning methods and select the most suitable model. The selection is made through the weightage from the users’ data to detect bone cancer. We show the model selected meets the expectations with the highest accuracy 90.36% using the residual neural network(ResNet101) algorithm and 89.51% precision in the prediction tasks.
Taxonomy of hybrid architectures involving rule-based reasoning and machine learning in clinical decision systems: A scoping review
2023, Journal of Biomedical Informatics
As the application of Artificial Intelligence (AI) technologies increases in the healthcare sector, the industry faces a need to combine medical knowledge, often expressed as clinical rules, with advances in machine learning (ML), which offer high prediction accuracy at the expense of transparency of decision making.
This paper seeks to review the present literature, identify hybrid architecture patterns that incorporate rules and machine learning, and evaluate the rationale behind their selection to inform future development and research on the design of transparent and precise clinical decision systems.
PubMed, IEEE Explore, and Google Scholar were queried in search for papers from 1992 to 2022, with the keywords: “clinical decision system”, “hybrid clinical architecture”, “machine learning and clinical rules”. Excluded articles did not use both ML and rules or did not provide any explanation of employed architecture. A proposed taxonomy was used to organize the results, analyze them, and depict them in graphical and tabular form. Two researchers, one with expertise in rule-based systems and another in ML, reviewed identified papers and discussed the work to minimize bias, and the third one re-reviewed the work to ensure consistency of reporting.
The authors screened 957 papers and reviewed 71 that met their criteria. Five distinct architecture archetypes were determined: Rules are Embedded in ML architecture (REML) (most used), ML pre-processes input data for Rule-Based inference (MLRB), Rule-Based method pre-processes input data for ML prediction (RBML), Rules influence ML training (RMLT), Parallel Ensemble of Rules and ML (PERML), which was rarely observed in clinical contexts.
Most architectures in the reviewed literature prioritize prediction accuracy over explainability and trustworthiness, which has led to more complex embedded approaches. Alternatively, parallel (PERML) architectures may be employed, allowing for a more transparent system that is easier to explain to patients and clinicians. The potential of this approach warrants further research.
A limitation of the study may be that it reviews scientific literature, while algorithms implemented in clinical practice may present different distributions of motivations and implementations of hybrid architectures.
Breast tumor localization and segmentation using machine learning techniques: Overview of datasets, findings, and methods
2023, Computers in Biology and Medicine
The Global Cancer Statistics 2020 reported breast cancer (BC) as the most common diagnosis of cancer type. Therefore, early detection of such type of cancer would reduce the risk of death from it. Breast imaging techniques are one of the most frequently used techniques to detect the position of cancerous cells or suspicious lesions. Computer-aided diagnosis (CAD) is a particular generation of computer systems that assist experts in detecting medical image abnormalities. In the last decades, CAD has applied deep learning (DL) and machine learning approaches to perform complex medical tasks in the computer vision area and improve the ability to make decisions for doctors and radiologists. The most popular and widely used technique of image processing in CAD systems is segmentation which consists of extracting the region of interest (ROI) through various techniques. This research provides a detailed description of the main categories of segmentation procedures which are classified into three classes: supervised, unsupervised, and DL. The main aim of this work is to provide an overview of each of these techniques and discuss their pros and cons. This will help researchers better understand these techniques and assist them in choosing the appropriate method for a given use case.
Improving decision making in the management of hospital readmissions using modern survival analysis techniques
2022, Decision Support Systems
Hospital readmissions lead to unnecessary demand for healthcare resources, greater financial costs, and poorer patient outcomes. These consequences have led hospitals to attempt to identify high-risk patients with predictive models, but research has rarely focused on survival analysis techniques, model applications, and performance measures. This study establishes the uses of survival models to support managerial decision-making for readmissions. First, machine learning and statistical survival techniques are applied, ten of which have not been used in previous readmission research. Secondly, applications of survival models in a decision support capacity are proposed, relating to intervention targeting, follow-up care customisation, and demand forecasting. Thirdly, performance measures for the proposed applications are determined and used for empirical model assessment. These performance measures have not been applied in previous readmission research. The empirical assessment is based on adult admissions to the Emergency Department of Gold Coast University Hospital (n = 46,659) and Robina Hospital (n = 23,976) in Queensland, Australia. The relevant aspects of performance were determined to be discrimination and calibration, as measured by time-dependent concordance and D-Calibration respectively. A range of discrimination and calibration combinations can be achieved by different models, with the Recursively Imputed Survival Tree, Cox regression, and hybrid Cox-ANN techniques being most promising. Survival approaches linking techniques, proposed applications, and performance measurement should be given greater consideration in future healthcare research and in institutions aiming to manage readmissions.
Quantitative sleep EEG synchronization analysis for automatic arousals detection
2020, Biomedical Signal Processing and Control
Electroencephalographic arousals are considered to be the main reason for the interruption of sleep and are visually examined by sleep physicians. Visual scoring of all-night recordings has inter-scorer variability which may lead to subjective results. Hence, we aimed to develop a novel automated method to detect arousals from two electroencephalographic channels in terms of the synchronic events of the right and left hemispheres.
In the context of the occurrence of arousal pattern, the relationship between two synchronic C3-A2 and C4-A1 channels were quantified using by coherence spectrum and mutual information. The power and the ratio values of the sub-bands of the coherence spectrum were selected as the five features. Furthermore, the mutual information value was determined as the sixth feature. The automatic detection performance was evaluated using six features and machine learning techniques, on five different patients' whole-night electroencephalography recordings. The presented method does not include any signal conditioning, pre-processing steps, any manual involvement, meta-rule-based approaches, and some empirical thresholds.
The significant increases were found in sub-bands of the coherence spectrum in case of arousal. Moreover, the mutual information of these channels was distinctive during the arousal state. Consequently, the overall accuracy, sensitivity, specificity, and PPV values were achieved as 99.5 %, 99.8 %, 99.6 %, and 99.3 %, respectively with using ensemble bagged tree.
The novelty of the present study is the practical determination of the relationship between electroencephalographic synchronization and the occurrence of the arousals between the central regions of the right and left hemispheres.
Practical Machine Learning for Data Analysis Using Python
2020, Practical Machine Learning for Data Analysis Using Python

View all citing articles on Scopus

View full text

A combined neural network and decision trees model for prognosis of breast cancer relapse

Abstract

Introduction

Section snippets

Breast cancer overview

Patient data

Results and discussion

Conclusions

Acknowledgements

Lancet

Neural Networks

Neural Networks

Ann. Oncol.

Comput Biomed Res

Artif Intell Med

Int J Biomed Comput

Artif Intell Med

J Mol Biol

Artif Intell Med

C-Net: a method for generating non-deterministic and dynamic multivariate decision trees

Know. Inform. Syst.

Statistical theory of overtraining—is cross-validation asymptotically effective?

Adv. Neural Inform. Process. Syst.

A further comparison of splitting rules for decision-tree induction

Mach. Learn.

Regression models and life tables

J. R. Stat. Soc.

Trio learning: a new strategy for building hybrid neural trees

Neural Syst.