Comparison of Fuzzy Inference, Logistic Regression, and Classification Trees (CART)

Machine-Learning Applications in Oral Cancer: A Systematic Review

Over the years, several machine-learning applications have been suggested to assist in various clinical scenarios relevant to oral cancer. We offer a systematic review to identify, assess, and summarize the evidence for reported uses in the areas of oral cancer detection and prevention, prognosis, pre-cancer, treatment, and quality of life. The main algorithms applied in the context of oral cancer applications corresponded to SVM, ANN, and LR, comprising 87.71% of the total published articles in the field. Genomic, histopathological, image, medical/clinical, spectral, and speech data were used most often to predict the four areas of application found in this review. In conclusion, our study has shown that machine-learning applications are useful for prognosis, diagnosis, and prevention of potentially malignant oral lesions (pre-cancer) and therapy. Nevertheless, we strongly recommended the application of these methods in daily clinical practice.

Prediction models applying machine learning to oral cavity cancer outcomes: A systematic review

OBJECTIVES

Machine learning platforms are now being introduced into modern oncological practice for classification and prediction of patient outcomes. To determine the current status of the application of these learning models as adjunctive decision-making tools in oral cavity cancer management, this systematic review aims to summarize the accuracy of machine-learning based models for disease outcomes.

METHODS

Electronic databases including PubMed, Scopus, EMBASE, Cochrane Library, LILACS, SciELO, PsychINFO, and Web of Science were searched up until December 21, 2020. Pertinent articles detailing the development and accuracy of machine learning prediction models for oral cavity cancer outcomes were selected in a two-stage process. Quality assessment was conducted using the Quality in Prognosis Studies (QUIPS) tool and results of base studies were qualitatively synthesized by all authors. Outcomes of interest were malignant transformation of precancer lesions, cervical lymph node metastasis, as well as treatment response, and prognosis of oral cavity cancer.

RESULTS

Twenty-seven articles out of 950 citations identified from electronic and manual searching were included in this study. Five studies had low bias concerns on the QUIPS tool. Prediction of malignant transformation, cervical lymph node metastasis, treatment response, and prognosis were reported in three, six, eight, and eleven articles respectively. Accuracy of these learning models on the internal or external validation sets ranged from 0.85 to 0.97 for malignant transformation prediction, 0.78-0.91 for cervical lymph node metastasis prediction, 0.64-1.00 for treatment response prediction, and 0.71-0.99 for prognosis prediction. In general, most trained algorithms predicting these outcomes performed better than alternate methods of prediction. We also found that models including molecular markers in training data had better accuracy estimates for malignant transformation, treatment response, and prognosis prediction.

CONCLUSION

Machine learning algorithms have a satisfactory to excellent accuracy for predicting three of four oral cavity cancer outcomes i.e., malignant transformation, nodal metastasis, and prognosis. However, considering the training approach of many available classifiers, these models may not be streamlined enough for clinical application currently.

Machine Learning Predicts Lymph Node Metastasis in Early-Stage Oral Tongue Squamous Cell Carcinoma

PURPOSE

Early-stage oral tongue squamous cell cancer (OTSCC) has a rate of metastasis to the cervical lymph nodes of 20 to 50%. This study aimed to build and validate 4 machine learning (ML) models to predict the occurrence of lymph node metastasis before and after surgery for early-stage (cT1N0 to cT2N0) OTSCC.

MATERIALS AND METHODS

We designed a retrospective cross-sectional study and reviewed the clinical and pathologic records of patients with early-stage OTSCC. The sample was composed of 2 groups with different node status (positive or negative) and was randomly split into training (70%) and testing (30%) sets. Four common ML algorithms-logistic regression, random forest, support vector machine, and naive Bayes-were used to predict pathologic nodal metastasis of early-stage OTSCC. The area under the receiver operating characteristic curve (AUC), sensitivity, and specificity were used to assess the performance of these models and conventional methods including depth of invasion (DOI), neutrophil-to-lymphocyte ratio (NLR), and tumor budding.

RESULTS

A total of 145 patients (56 with positive and 89 with negative lymph nodes) were included in this study. The performance of ML models was significantly superior to that of conventional prediction methods. The random forest model performed best (AUC, 0.786; sensitivity, 85%; specificity, 75%) and exceeded the performance of NLR (AUC, 0.539; sensitivity, 53.6%; specificity, 53.9%; P = .003). When DOI, worst pattern of invasion, lymphocytic host response, and tumor budding were added to model analysis according to patients' postoperative pathologic records, the support vector machine model performed best (AUC, 0.956; sensitivity, 100%; specificity, 87.5%) and was superior to univariate assessment of tumor budding (AUC, 0.830; sensitivity, 80.9%; specificity, 87.5%, P = .002), DOI (AUC, 0.613; sensitivity, 91.1%; specificity, 31.5%; P < .001), and NLR.

CONCLUSIONS

ML shows a better performance in predicting lymph node metastasis of early-stage OTSCC than conventional prediction methods of DOI, NLR, or tumor budding.

Clinical characteristics and disease-specific prognostic nomogram for primary gliosarcoma: a SEER population-based analysis

Because the study population with gliosarcoma (GSM) is limited, the understanding of this disease is insufficient. In this study, the authors aimed to determine the clinical characteristics and independent prognostic factors influencing the prognosis of GSM patients and to develop a nomogram to predict the prognosis of GSM patients after craniotomy. A total of 498 patients diagnosed with primary GSM between 2004 and 2015 were extracted from the 18 Registries Research Data of the Surveillance, Epidemiology, and End Results (SEER) database. The median disease-specific survival (DSS) was 12.0 months, and the postoperative 0.5-, 1-, and 3-year DSS rates were 71.4%, 46.4% and 9.8%, respectively. We applied both the Cox proportional hazards model and the decision tree model to determine the prognostic factors of primary GSM. The Cox proportional hazards model demonstrated that age at presentation, tumour size, metastasis state and adjuvant chemotherapy (CT) were independent prognostic factors for DSS. The decision tree model suggested that age <71 years and adjuvant CT were associated with a better prognosis for GSM patients. The nomogram generated via the Cox proportional hazards model was developed by applying the rms package in R version 3.5.0. The C-index of internal validation for DSS prediction was 0.67 (95% confidence interval (CI), 0.63 to 0.70). The calibration curve at one year suggested that there was good consistency between the predicted DSS and the actual DSS probability. This study was the first to develop a disease-specific nomogram for predicting the prognosis of primary GSM patients after craniotomy, which can help clinicians immediately and accurately predict patient prognosis and conduct further treatment.

To Explore Intracerebral Hematoma with a Hybrid Approach and Combination of Discriminative Factors

OBJECTIVES

To find discriminative combination of influential factors of Intracerebral hematoma (ICH) to cluster ICH patients with similar features to explore relationship among influential factors and 30-day mortality of ICH.

METHODS

The data of ICH patients are collected. We use a decision tree to find discriminative combination of the influential factors. We cluster ICH patients with similar features using Fuzzy C-means algorithm (FCM) to construct a support vector machine (SVM) for each cluster to build a multi-SVM classifier. Finally, we designate each testing data into its appropriate cluster and apply the corresponding SVM classifier of the cluster to explore the relationship among impact factors and 30-day mortality.

RESULTS

The two influential factors chosen to split the decision tree are Glasgow coma scale (GCS) score and Hematoma size. FCM algorithm finds three centroids, one for high danger group, one for middle danger group, and the other for low danger group. The proposed approach outperforms benchmark experiments without FCM algorithm to cluster training data.

CONCLUSIONS

It is appropriate to construct a classifier for each cluster with similar features. The combination of factors with significant discrimination as input variables should outperform that with only single discriminative factor as input variable.

Which is a more accurate predictor in colorectal survival analysis? Nine data mining algorithms vs. the TNM staging system

OBJECTIVE

Over the past decades, many studies have used data mining technology to predict the 5-year survival rate of colorectal cancer, but there have been few reports that compared multiple data mining algorithms to the TNM classification of malignant tumors (TNM) staging system using a dataset in which the training and testing data were from different sources. Here we compared nine data mining algorithms to the TNM staging system for colorectal survival analysis.

METHODS

Two different datasets were used: 1) the National Cancer Institute's Surveillance, Epidemiology, and End Results dataset; and 2) the dataset from a single Chinese institution. An optimization and prediction system based on nine data mining algorithms as well as two variable selection methods was implemented. The TNM staging system was based on the 7(th) edition of the American Joint Committee on Cancer TNM staging system.

RESULTS

When the training and testing data were from the same sources, all algorithms had slight advantages over the TNM staging system in predictive accuracy. When the data were from different sources, only four algorithms (logistic regression, general regression neural network, bayesian networks, and Naïve Bayes) had slight advantages over the TNM staging system. Also, there was no significant differences among all the algorithms (p>0.05).

CONCLUSIONS

The TNM staging system is simple and practical at present, and data mining methods are not accurate enough to replace the TNM staging system for colorectal cancer survival prediction. Furthermore, there were no significant differences in the predictive accuracy of all the algorithms when the data were from different sources. Building a larger dataset that includes more variables may be important for furthering predictive accuracy.

Artificial neural networks and decision tree model analysis of liver cancer proteomes

Hepatocellular carcinoma (HCC) is a heterogeneous cancer and usually diagnosed at late advanced tumor stages of high lethality. The present study attempted to obtain a proteome-wide analysis of HCC in comparison with adjacent non-tumor liver tissues, in order to facilitate biomarkers' discovery and to investigate the mechanisms of HCC development. A cohort of 66 Chinese patients with HCC was included for proteomic profiling study by two-dimensional gel electrophoresis (2-DE) analysis. Artificial neural network (ANN) and decision tree (CART) data-mining methods were employed to analyze the profiling data and to delineate significant patterns and trends for discriminating HCC from non-malignant liver tissues. Protein markers were identified by tandem MS/MS. A total of 132 proteome datasets were generated by 2-DE expression profiling analysis, and each with 230 consolidated protein expression intensities. Both the data-mining algorithms successfully distinguished the HCC phenotype from other non-malignant liver samples. The detection sensitivity and specificity of ANN were 96.97% and 87.88%, while those of CART were 81.82% and 78.79%, respectively. The three biological classifiers in the CART model were identified as cytochrome b5, heat shock 70 kDa protein 8 isoform 2, and cathepsin B. The 2-DE-based proteomic profiling approach combined with the ANN or CART algorithm yielded satisfactory performance on identifying HCC and revealed potential candidate cancer biomarkers.

A comparison of regression trees, logistic regression, generalized additive models, and multivariate adaptive regression splines for predicting AMI mortality

Clinicians and health service researchers are frequently interested in predicting patient-specific probabilities of adverse events (e.g. death, disease recurrence, post-operative complications, hospital readmission). There is an increasing interest in the use of classification and regression trees (CART) for predicting outcomes in clinical studies. We compared the predictive accuracy of logistic regression with that of regression trees for predicting mortality after hospitalization with an acute myocardial infarction (AMI). We also examined the predictive ability of two other types of data-driven models: generalized additive models (GAMs) and multivariate adaptive regression splines (MARS). We used data on 9484 patients admitted to hospital with an AMI in Ontario. We used repeated split-sample validation: the data were randomly divided into derivation and validation samples. Predictive models were estimated using the derivation sample and the predictive accuracy of the resultant model was assessed using the area under the receiver operating characteristic (ROC) curve in the validation sample. This process was repeated 1000 times-the initial data set was randomly divided into derivation and validation samples 1000 times, and the predictive accuracy of each method was assessed each time. The mean ROC curve area for the regression tree models in the 1000 derivation samples was 0.762, while the mean ROC curve area of a simple logistic regression model was 0.845. The mean ROC curve areas for the other methods ranged from a low of 0.831 to a high of 0.851. Our study shows that regression trees do not perform as well as logistic regression for predicting mortality following AMI. However, the logistic regression model had performance comparable to that of more flexible, data-driven models such as GAMs and MARS.

Bayesian analysis, pattern analysis, and data mining in health care

PURPOSE OF REVIEW

To discuss the current role of data mining and Bayesian methods in biomedicine and heath care, in particular critical care.

RECENT FINDINGS

Bayesian networks and other probabilistic graphical models are beginning to emerge as methods for discovering patterns in biomedical data and also as a basis for the representation of the uncertainties underlying clinical decision-making. At the same time, techniques from machine learning are being used to solve biomedical and health-care problems.

SUMMARY

With the increasing availability of biomedical and health-care data with a wide range of characteristics there is an increasing need to use methods which allow modeling the uncertainties that come with the problem, are capable of dealing with missing data, allow integrating data from various sources, explicitly indicate statistical dependence and independence, and allow integrating biomedical and clinical background knowledge. These requirements have given rise to an influx of new methods into the field of data analysis in health care, in particular from the fields of machine learning and probabilistic graphical models.