Comparison of prediction models for adverse outcome in pediatric meningococcal disease using artificial neural network and logistic regression analyses

Prediction of gastrointestinal cancers in the ONCONUT cohort study: comparison between logistic regression and artificial neural network

Background

Artificial neural networks (ANNs) and logistic regression (LR) are the models of chosen in many medical data classification tasks. Several published articles were based on summarizing the differences and similarities of these models from a technical point of view and critically assessing the quality of the models. The aim of this study was to compare ANN and LR the statistical techniques to predict gastrointestinal cancer in an elderly cohort in Southern Italy (ONCONUT study).

Method

In 1992, ONCONUT was started with the aim of evaluating the relationship between diet and cancer development in a Southern Italian elderly population. Patients with gastrointestinal cancer (ICD-10 from 150.0 to 159.9) were included in the study (n = 3,545).

Results

This cohort was used to train and test the ANN and LR. LR was evaluated separately for macro- and micronutrients, and the accuracy was evaluated based on true positives and true negatives versus the total (97.15%). Then, ANN was trained and the accuracy was evaluated (96.61% for macronutrients and 97.06% for micronutrients). To further investigate the classification capabilities of ANN, k-fold cross-validation and genetic algorithm (GA) were used after balancing the dataset among classes.

Conclusions

Both LR and ANN had high accuracy and similar performance. Both models had the potential to be used as decision clinical support integrated into clinical practice, because in many circumstances, the use of a simple LR model was likely to be adequate for real-world needs, but in others in which there were large amounts of data, the application of advanced analytic tools such as ANNs could be indicated, and the GA optimizer needed to optimize the accuracy of ANN.

Classification Performance of Neural Networks Versus Logistic Regression Models: Evidence From Healthcare Practice

Machine learning encompasses statistical approaches such as logistic regression (LR) through to more computationally complex models such as neural networks (NN). The aim of this study is to review current published evidence for performance from studies directly comparing logistic regression, and neural network classification approaches in medicine.

A literature review was carried out to identify primary research studies which provided information regarding comparative area under the curve (AUC) values for the overall performance of both LR and NN for a defined clinical healthcare-related problem. Following an initial search, articles were reviewed to remove those that did not meet the criteria and performance metrics were extracted from the included articles. Teh initial search revealed 114 articles; 21 studies were included in the study. In 13/21 (62%) of cases, NN had a greater AUC compared to LR, but in most the difference was small and unlikely to be of clinical significance; (unweighted mean difference in AUC 0.03 (95% CI 0-0.06) in favour of NN versus LR.

In the majority of cases examined across a range of clinical settings, LR models provide reasonable performance that is only marginally improved using more complex methods such as NN. In many circumstances, the use of a relatively simple LR model is likely to be adequate for real-world needs but in specific circumstances in which large amounts of data are available, and where even small increases in performance would provide significant management value, the application of advanced analytic tools such as NNs may be indicated.

Predictive Values of Preoperative Characteristics for 30-Day Mortality in Traumatic Hip Fracture Patients

Hip fracture patients have a high risk of mortality after surgery, with 30-day postoperative rates as high as 10%. This study aimed to explore the predictive ability of preoperative characteristics in traumatic hip fracture patients as they relate to 30-day postoperative mortality using readily available variables in clinical practice. All adult patients who underwent primary emergency hip fracture surgery in Sweden between 2008 and 2017 were included in the analysis. Associations between the possible predictors and 30-day mortality was performed using a multivariate logistic regression (LR) model; the bidirectional stepwise method was used for variable selection. An LR model and convolutional neural network (CNN) were then fitted for prediction. The relative importance of individual predictors was evaluated using the permutation importance and Gini importance. A total of 134,915 traumatic hip fracture patients were included in the study. The CNN and LR models displayed an acceptable predictive ability for predicting 30-day postoperative mortality using a test dataset, displaying an area under the ROC curve (AUC) of as high as 0.76. The variables with the highest importance in prediction were age, sex, hypertension, dementia, American Society of Anesthesiologists (ASA) classification, and the Revised Cardiac Risk Index (RCRI). Both the CNN and LR models achieved an acceptable performance in identifying patients at risk of mortality 30 days after hip fracture surgery. The most important variables for prediction, based on the variables used in the current study are age, hypertension, dementia, sex, ASA classification, and RCRI.

Comparison of new modeling methods for postnatal weight in ELBW infants using prenatal and postnatal data

OBJECTIVES

Postnatal infant weight curves are used to assess fluid management and evaluate postnatal nutrition and growth. Traditionally, postnatal weight curves are based on birth weight and do not incorporate postnatal clinical information. The aim of the present study was to compare the accuracy of birth weight-based weight curves with weight curves created from individual patient records, including electronic records, using 2 predictive modeling methods, linear regression (LR) and an artificial neural network (NN), which apply mathematical relations between predictor and outcome variables.

METHODS

Perinatal demographic and postnatal nutrition data were collected for extremely-low-birth-weight (ELBW; birth weight <1000 g) infants. Static weight curves were generated using published algorithms. The postnatal predictive models were created using the demographic and nutrition dataset.

RESULTS

Birth weight (861 ± 83 g, mean ± 1 standard deviation [SD]), gestational age (26.2 ± 1.4 weeks), and the first month of nutrition data were collected from individual health records for 92 ELBW infants. The absolute residual (|measured-predicted|) for weight was 84.8 ± 74.4 g for the static weight curves, 60.9 ± 49.1 g for the LR model, and 12.9 ± 9.2 g for the NN model, analysis of variance: both LR and NN P<0.01 versus static curve. NPO (nothing by mouth) infants had greater weight curve discrepancies.

CONCLUSIONS

Compared with birth weight-based and logistic regression-generated weight curves, NN-generated weight curves more closely approximated ELBW infant weight curves, and, using the present electronic health record systems, may produce weight curves better reflective of the patient's status.

Evaluation of two treatment outcome prediction models for restoration of visual fields in patients with postchiasmatic visual pathway lesions

Visual functions of patients with visual field defects after acquired brain injury affecting the primary visual pathway can be improved by means of vision restoration training. Since the extent of the restored visual field varies between patients, the prediction of treatment outcome and its visualization may help patients to decide for or against participating in therapies aimed at vision restoration. For this purpose, two treatment outcome prediction models were established based on either self-organizing maps (SOMs) or categorical regression (CR) to predict visual field change after intervention by several features that were hypothesized to be associated with vision restoration. Prediction was calculated for visual field changes recorded with High Resolution Perimetry (HRP). Both models revealed a similar predictive quality with the CR model being slightly more beneficial. Predictive quality of the SOM model improved when using only a small number of features that exhibited a higher association with treatment outcome than the remaining features, i.e. neighborhood activity and homogeneity within the surrounding 5° visual field of a given position, together with its residual function and distance to the scotoma border. Although both models serve their purpose, these were not able to outperform a primitive prediction rule that attests the importance of areas of residual vision, i.e. regions with partial visual field function, for vision restoration.

Mortality predicted accuracy for hepatocellular carcinoma patients with hepatic resection using artificial neural network

The aim of this present study is firstly to compare significant predictors of mortality for hepatocellular carcinoma (HCC) patients undergoing resection between artificial neural network (ANN) and logistic regression (LR) models and secondly to evaluate the predictive accuracy of ANN and LR in different survival year estimation models. We constructed a prognostic model for 434 patients with 21 potential input variables by Cox regression model. Model performance was measured by numbers of significant predictors and predictive accuracy. The results indicated that ANN had double to triple numbers of significant predictors at 1-, 3-, and 5-year survival models as compared with LR models. Scores of accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUROC) of 1-, 3-, and 5-year survival estimation models using ANN were superior to those of LR in all the training sets and most of the validation sets. The study demonstrated that ANN not only had a great number of predictors of mortality variables but also provided accurate prediction, as compared with conventional methods. It is suggested that physicians consider using data mining methods as supplemental tools for clinical decision-making and prognostic evaluation.

Disease-free survival after hepatic resection in hepatocellular carcinoma patients: a prediction approach using artificial neural network

Background

A database for hepatocellular carcinoma (HCC) patients who had received hepatic resection was used to develop prediction models for 1-, 3- and 5-year disease-free survival based on a set of clinical parameters for this patient group.

Methods

The three prediction models included an artificial neural network (ANN) model, a logistic regression (LR) model, and a decision tree (DT) model. Data for 427, 354 and 297 HCC patients with histories of 1-, 3- and 5-year disease-free survival after hepatic resection, respectively, were extracted from the HCC patient database. From each of the three groups, 80% of the cases (342, 283 and 238 cases of 1-, 3- and 5-year disease-free survival, respectively) were selected to provide training data for the prediction models. The remaining 20% of cases in each group (85, 71 and 59 cases in the three respective groups) were assigned to validation groups for performance comparisons of the three models. Area under receiver operating characteristics curve (AUROC) was used as the performance index for evaluating the three models.

Conclusions

The ANN model outperformed the LR and DT models in terms of prediction accuracy. This study demonstrated the feasibility of using ANNs in medical decision support systems for predicting disease-free survival based on clinical databases in HCC patients who have received hepatic resection.

Predicting three-year kidney graft survival in recipients with systemic lupus erythematosus

Predicting the outcome of kidney transplantation is important in optimizing transplantation parameters and modifying factors related to the recipient, donor, and transplant procedure. As patients with end-stage renal disease (ESRD) secondary to lupus nephropathy are generally younger than the typical ESRD patients and also seem to have inferior transplant outcome, developing an outcome prediction model in this patient category has high clinical relevance. The goal of this study was to compare methods of building prediction models of kidney transplant outcome that potentially can be useful for clinical decision support. We applied three well-known data mining methods (classification trees, logistic regression, and artificial neural networks) to the data describing recipients with systemic lupus erythematosus (SLE) in the US Renal Data System (USRDS) database. The 95% confidence interval (CI) of the area under the receiver-operator characteristic curves (AUC) was used to measure the discrimination ability of the prediction models. Two groups of predictors were selected to build the prediction models. Using input variables based on Weka (a open source machine learning software) supplemented with additional variables of known clinical relevance (38 total predictors), the logistic regression performed the best overall (AUC: 0.74, 95% CI: 0.72-0.77)-significantly better (p < 0.05) than the classification trees (AUC: 0.70, 95% CI: 0.67-0.72) but not significantly better (p = 0.218) than the artificial neural networks (AUC: 0.71, 95% CI: 0.69-0.73). The performance of the artificial neural networks was not significantly better than that of the classification trees (p = 0.693). Using the more parsimonious subset of variables (six variables), the logistic regression (AUC: 0.73, 95% CI: 0.71-0.75) did not perform significantly better than either the classification tree (AUC: 0.70, 95% CI: 0.68-0.73) or the artificial neural network (AUC: 0.73, 95% CI: 0.70-0.75) models. We generated several models predicting 3-year allograft survival in kidney transplant recipients with SLE that potentially can be used in practice. The performance of logistic regression and classification tree was not inferior to more complex artificial neural network. Prediction models may be used in clinical practice to identify patients at risk.

Determination of the early time of death by computerized image analysis of DNA degradation: which is the best quantitative indicator of DNA degradation?

This study evaluated the correlation between DNA degradation of the splenic lymphocytes and the early time of death, examined the early time of death by computerized image analysis technique (CIAT) and identified the best parameter that quantitatively reflects the DNA degradation. The spleen tissues from 34 SD rats were collected, subjected to cell smearing every 2 h within the first 36 h after death, stained by Feulgen-Van's staining, three indices reflecting DNA content in splenic lymphocytes, including integral optical density (IOD), average optical density (AOD), average gray scale (AG) were measured by the image analysis. Our results showed that IOD and AOD decreased and AG increased over time within the first 36 h. A stepwise linear regression analysis showed that only AG was fitted. A correlation between the postmortem interval (PMI) and AG was identified and the corresponding regression equation was obtained. Our study suggests that CIAT is a useful and promising tool for the estimation of early PMI with good objectivity and reproducibility, and AG is a more effective and better quantitative indicator for the estimation of PMI within the first 36 h after death in rats.

Comparison among probabilistic neural network, support vector machine and logistic regression for evaluating the effect of subthalamic stimulation in Parkinson disease on ground reaction force during gait

Deep brain stimulation of the subthalamic nucleus (DBS-STN) is an approved treatment for advanced Parkinson disease (PD) patients; however, there is a need to further evaluate its effect on gait. This study compares logistic regression (LR), probabilistic neural network (PNN) and support vector machine (SVM) classifiers for discriminating between normal and PD subjects in assessing the effects of DBS-STN on ground reaction force (GRF) with and without medication. Gait analysis of 45 subjects (30 normal and 15 PD subjects who underwent bilateral DBS-STN) was performed. PD subjects were assessed under four test conditions: without treatment (mof-sof), with stimulation alone (mof-son), with medication alone (mon-sof), and with medication and stimulation (mon-son). Principal component (PC) analysis was applied to the three components of GRF separately, where six PC scores from vertical, one from anterior-posterior and one from medial-lateral were chosen by the broken stick test. Stepwise LR analysis employed the first two and fifth vertical PC scores as input variables. Using the bootstrap approach to compare model performances for classifying GRF patterns from normal and untreated PD subjects, the first three and the fifth vertical PCs were attained as SVM input variables, while the same ones plus the first anterior-posterior were selected as PNN input variables. PNN performed better than LR and SVM according to area under the receiver operating characteristic curve and the negative likelihood ratio. When evaluating treatment effects, the classifiers indicated that DBS-STN alone was more effective than medication alone, but the greatest improvements occurred with both treatments together.

Nonlinear association between serum testosterone levels and coronary artery disease in Iranian men

Previous studies have shown controversial results about the role of androgens in coronary artery disease (CAD). We performed this study to examine and compare the relationship between androgenic hormones and CAD using conventional linear statistical techniques as well as novel non-linear approaches. The study was conducted on 502 consecutive men who were referred for selective coronary angiography at Tehran Heart Center due to different indications. We studied the relationship between androgenic hormones and CAD by using the generalized linear models, generalized additive models, and neural networks. Free testosterone (fT), total testosterone (tT) and dehydroepiandrosterone sulfate levels in patients with significant CAD versus normal individuals were 6.69 +/- 3.20 pg/ml, 16.60 +/- 6.66 nm/l, and 113.38 +/- 72.9 microg/dl versus 7.12 +/- 3.58 pg/ml, 15.82 +/- 7.26 nm/l, and 109.03 +/- 68.19 microg/dl, respectively (P > 0.05). The Generalized linear models was unable to show any significant relationship between androgenic hormones and CAD, while generalized additive model and neural networks supported the significant effect of androgenic hormones on CAD. This finding suggests a nonlinear association of tT levels with CAD: lower levels have a preventive effect on CAD, whereas higher values increase the risk of CAD. Emphasizing the non-linearity of the variables may provide new insight into the possible explanation of the effect of androgenic hormones on CAD.

Prediction of periventricular leukomalacia. Part II: Selection of hemodynamic features using computational intelligence

OBJECTIVE

The objective of Part II is to analyze the dataset of extracted hemodynamic features (Case 3 of Part I) through computational intelligence (CI) techniques for identification of potential prognostic factors for periventricular leukomalacia (PVL) occurrence in neonates with congenital heart disease.

METHODS

The extracted features (Case 3 dataset of Part I) were used as inputs to CI based classifiers, namely, multi-layer perceptron (MLP) and probabilistic neural network (PNN) in combination with genetic algorithms (GA) for selection of the most suitable features predicting the occurrence of PVL. The selected features were next used as inputs to a decision tree (DT) algorithm for generating easily interpretable rules of PVL prediction.

RESULTS

Prediction performance for two CI based classifiers, MLP and PNN coupled with GA are presented for different number of selected features. The best prediction performances were achieved with 6 and 7 selected features. The prediction success was 100% in training and the best ranges of sensitivity (SN), specificity (SP) and accuracy (AC) in test were 60-73%, 74-84% and 71-74%, respectively. The identified features when used with the DT algorithm gave best SN, SP and AC in the ranges of 87-90% in training and 80-87%, 74-79% and 79-82% in test. Among the variables selected in CI, systolic and diastolic blood pressures, and pCO(2) figured prominently similar to Part I. Decision tree based rules for prediction of PVL occurrence were obtained using the CI selected features.

CONCLUSIONS

The proposed approach combines the generalization capability of CI based feature selection approach and generation of easily interpretable classification rules of the decision tree. The combination of CI techniques with DT gave substantially better test prediction performance than using CI and DT separately.

Developing a prognostic model for 90-day mortality after liver transplantation based on pretransplant recipient factors

BACKGROUND

Current statistical prognostic models for mortality after liver transplantation do not have good discriminatory ability. Furthermore, the methodology used to develop these models is often flawed. The objective of this paper is to develop a prognostic model for 90-day mortality after liver transplantation based on pretransplant recipient factors, employing a rigorous model development method.

METHODS

We used data on 4,829 patient that were prospectively collected for the UK & Ireland Liver Transplant Audit. Switching regression was employed to impute missing values combined with a bootstrapping approach for variable selection.

RESULTS

In all, 452 patients (9.4%) died within 90 days of their transplantation. The final prognostic model was well calibrated and discriminated moderately well between patients who did and who did not die (c-statistic 0.65, 95% CI [0.63, 0.68]). Although discrimination was not excellent overall, the results showed that those patients with a "low" chance of dying within 90 days of their transplant and those with a "high" chance of dying could be differentiated from patients with a "intermediate" chance.

CONCLUSIONS

Our model can provide transplant candidates with predictions of their early posttransplantation prospects before any donor information is known, which is essential information for patients with end-stage liver disease for whom liver transplantation is a treatment option.

Wavelet neural network classification of EEG signals by using AR model with MLE preprocessing

Since EEG is one of the most important sources of information in therapy of epilepsy, several researchers tried to address the issue of decision support for such a data. In this paper, we introduce two fundamentally different approaches for designing classification models (classifiers); the traditional statistical method based on logistic regression and the emerging computationally powerful techniques based on artificial neural networks (ANNs). Logistic regression as well as feedforward error backpropagation artificial neural networks (FEBANN) and wavelet neural networks (WNN) based classifiers were developed and compared in relation to their accuracy in classification of EEG signals. In these methods we used FFT and autoregressive (AR) model by using maximum likelihood estimation (MLE) of EEG signals as an input to classification system with two discrete outputs: epileptic seizure or nonepileptic seizure. By identifying features in the signal we want to provide an automatic system that will support a physician in the diagnosing process. By applying AR with MLE in connection with WNN, we obtained novel and reliable classifier architecture. The network is constructed by the error backpropagation neural network using Morlet mother wavelet basic function as node activation function. The comparisons between the developed classifiers were primarily based on analysis of the receiver operating characteristic (ROC) curves as well as a number of scalar performance measures pertaining to the classification. The WNN-based classifier outperformed the FEBANN and logistic regression based counterpart. Within the same group, the WNN-based classifier was more accurate than the FEBANN-based classifier, and the logistic regression-based classifier.

Systematic review and validation of prognostic models in liver transplantation

A model that can accurately predict post-liver transplant mortality would be useful for clinical decision making, would help to provide patients with prognostic information, and would facilitate fair comparisons of surgical performance between transplant units. A systematic review of the literature was carried out to assess the quality of the studies that developed and validated prognostic models for mortality after liver transplantation and to validate existing models in a large data set of patients transplanted in the United Kingdom (UK) and Ireland between March 1994 and September 2003. Five prognostic model papers were identified. The quality of the development and validation of all prognostic models was suboptimal according to an explicit assessment tool of the internal, external, and statistical validity, model evaluation, and practicality. The discriminatory ability of the identified models in the UK and Ireland data set was poor (area under the receiver operating characteristic curve always smaller than 0.7 for adult populations). Due to the poor quality of the reporting, the methodology used for the development of the model could not always be determined. In conclusion, these findings demonstrate that currently available prognostic models of mortality after liver transplantation can have only a limited role in clinical practice, audit, and research.

Comparative analysis of logistic regression and artificial neural network for computer-aided diagnosis of breast masses

RATIONALE AND OBJECTIVE

To compare logistic regression and artificial neural network for computer-aided diagnosis on breast sonograms.

MATERIALS AND METHODS

Ultrasound images of 24 malignant and 30 benign masses were analyzed quantitatively for margin sharpness, margin echogenicity, and angular variation in margin. These features and age of patients were used with two pattern classifiers, logistic regression, and an artificial neural network to differentiate between malignant and benign masses. The performance of two methods was compared by receiver operating characteristic (ROC) analysis.

RESULTS

The area under the ROC curve Az (+/-SD) of the logistic regression analysis was 0.853 +/- 0.059 with 95% confidence limit (0.760-0.950). The area under the ROC curve of the artificial neural network analysis was 0.856 +/- 0.058 with 95% confidence limit (0.734-0.936). Although both the logistic regression and the artificial neural network had the same area under the ROC curve, the shapes of two curves were different. At 95% sensitivity, the artificial neural network had 76.5% specificity, whereas logistic regression had 64.7% specificity.

CONCLUSION

There was no difference in performance between logistic regression and the artificial neural network as measured by the area under the ROC curve. However, at a fixed 95% sensitivity, the artificial neural network had higher (12%) specificity compared with logistic regression value.

Comparison of artificial neural network and logistic regression models for prediction of mortality in head trauma based on initial clinical data

Background

In recent years, outcome prediction models using artificial neural network and multivariable logistic regression analysis have been developed in many areas of health care research. Both these methods have advantages and disadvantages. In this study we have compared the performance of artificial neural network and multivariable logistic regression models, in prediction of outcomes in head trauma and studied the reproducibility of the findings.

Methods

1000 Logistic regression and ANN models based on initial clinical data related to the GCS, tracheal intubation status, age, systolic blood pressure, respiratory rate, pulse rate, injury severity score and the outcome of 1271 mainly head injured patients were compared in this study. For each of one thousand pairs of ANN and logistic models, the area under the receiver operating characteristic (ROC) curves, Hosmer-Lemeshow (HL) statistics and accuracy rate were calculated and compared using paired T-tests.

Results

ANN significantly outperformed logistic models in both fields of discrimination and calibration but under performed in accuracy. In 77.8% of cases the area under the ROC curves and in 56.4% of cases the HL statistics for the neural network model were superior to that for the logistic model. In 68% of cases the accuracy of the logistic model was superior to the neural network model.

Conclusions

ANN significantly outperformed the logistic models in both fields of discrimination and calibration but lagged behind in accuracy. This study clearly showed that any single comparison between these two models might not reliably represent the true end results. External validation of the designed models, using larger databases with different rates of outcomes is necessary to get an accurate measure of performance outside the development population.

Predicting extubation outcome in preterm newborns: a comparison of neural networks with clinical expertise and statistical modeling

Even though ventilator technology and monitoring of premature infants has improved immensely over the past decades, there are still no standards for weaning and determining optimal extubation time for those infants. Approximately 30% of intubated preterm infants will fail attempted extubation, requiring reintubation and resuming of mechanical ventilation. A machine-learning approach using artificial neural networks (ANNs) to aid in extubation decision making is hereby proposed. Using expert opinion, 51 variables were identified as being relevant for the decision of whether to extubate an infant who is on mechanical ventilation. The data on 183 premature infants, born between 1999 and 2002, were collected by review of medical charts. The ANN extubation model was compared with alternative statistical modeling using multivariate logistic regression and also with the clinician's own predictive insight using sensitivity analysis and receiver operating characteristic curves. The optimal ANN model used 13 parameters and achieved an area under the receiver operating characteristic curve of 0.87 (out-of-sample validation), comparing favorably with multivariate logistic regression. It also compared well with the clinician's expertise, which raises the possibility of being useful as an automated alert tool. Because an ANN learns directly from previous data obtained in the institution where it is to be used, this makes it particularly amenable for application to evidence-based medicine. Given the variety of practices and equipment being used in different hospitals, this may be particularly relevant in the context of caring for preterm newborns who are on mechanical ventilation.

Application of artificial neural network modelling to identify severely ill patients whose aminoglycoside concentrations are likely to fall below therapeutic concentrations

OBJECTIVE

Identification of ICU patients whose concentrations are likely to fall below therapeutic concentrations using artificial neural network (ANN) modelling and individual patient physiologic data.

METHOD

Data on indicators of disease severity and some physiologic data were collected from 89 ICU patients who received arbekacin (ABK) and 61 who received amikacin (AMK). Three-layer ANN modelling and multivariate logistic regression analysis were used to predict the plasma concentrations of the aminoglycosides (ABK and AMK) in the severely ill patients.

RESULTS

Predictive performance analysis showed that the sensitivity and specificity of ANN modelling was superior to multivariate logistic regression analysis. For accurate modelling, a predictable range should be inferred from the data structure before the analysis. Restriction of the predictable region, based on the data structure, increased predictive performance.

CONCLUSION

ANN analysis was superior to multivariate logistic regression analysis in predicting which patients would have plasma concentrations lower than the minimum therapeutic concentration. To improve predictive performance, the predictable range should be inferred from the data structure before prediction. When applying ANN modelling in clinical settings, the predictive performance and predictable region should be investigated in detail to avoid the risk of harm to severely ill patients.

Clinical application of artificial neural network (ANN) modeling to predict pharmacokinetic parameters of severely ill patients

Artificial neural network (ANN) modeling was used to evaluate the pharmacokinetics of aminoglycosides (arbekacin sulfate and amikacin sulfate) in severely ill patients. The plasma level was predicted by ANN modeling using parameters related to the severity of the patient's condition and the predictive performance was shown to be better than could be achieved using multiple regression analysis. These results indicate that there is a non-linear relationship between the pharmacokinetics of aminoglycosides and the severity of the patient's condition, and this should be taken into account when determining the dose for severely ill patients. Patients whose plasma levels are likely to fall below the effective level can be identified by ANN modeling with a predictive sensitivity and specificity superior to multivariate logistic regression analysis. The predictable range should be inferred from the data structure before the modeling in order to improve the predictive performance. The volume of distribution (Vd) in the normal range was weakly predicted by ANN modeling from the patients' data. Prediction of clearance by ANN modeling was poorer than that obtained from serum creatinine concentration by linear regression analysis. These results suggest that the input-output relationship (linear or non-linear) should be taken into account in selecting the modeling method. Linear modeling can give better predictive performance for linear systems and non-linear modeling can give better predictive performance for non-linear systems. In general, the performance of ANN modeling was superior to linear modeling for PK/PD prediction. For accurate modeling, a predictable range should be inferred from the data structure before the analysis. Restriction of the predictable region, as determined from the data structure, produced an increase in prediction performance. When applying ANN modeling in clinical settings, the predictive performance and predictable region should be investigated in detail to avoid the risk of harm to severely ill patients.