Predicting outcomes after liver transplantation. A connectionist approach

Artificial neural network for prediction of acute kidney injury after liver transplantation for cirrhosis and hepatocellular carcinoma

Acute kidney injury (AKI) has serious consequences on the prognosis of patients undergoing liver transplantation (LT) for liver cancer and cirrhosis. Artificial neural network (ANN) has recently been proposed as a useful tool in many fields in the setting of solid organ transplantation and surgical oncology, where patient prognosis depends on a multidimensional and nonlinear relationship between variables pertaining to the surgical procedure, the donor (graft characteristics), and the recipient comorbidities. In the specific case of LT, ANN models have been developed mainly to predict survival in patients with cirrhosis, to assess the best donor-to-recipient match during allocation processes, and to foresee postoperative complications and outcomes. This is a specific opinion review on the role of ANN in the prediction of AKI after LT for liver cancer and cirrhosis, highlighting potential strengths of the method to forecast this serious postoperative complication.

The liver transplant risk score prognosticates the outcomes of liver transplant recipients at listing

BACKGROUND

We assessed if the risk of post-liver transplant mortality within 24 h could be stratified at the time of listing using the liver transplant risk score (LTRS). Secondary aims were to assess if the LTRS could stratify the risk of 30-day, 1-year mortality, and survival beyond the first year.

METHODS

MELD, BMI, age, diabetes, and the need for dialysis were the five variables used to calculate the LTRS during patients' evaluation for liver transplantation. Mortality rates at 24 h, 30 days, and 1-year were compared among groups of patients with different LTRS. Patients with ABO-incompatibility, redo, multivisceral, partial graft and malignancies except for hepatocellular carcinoma were excluded. Data of 48,616 adult liver transplant recipients were extracted from the Scientific Registry of Transplant Recipients between 2002 and 2017.

RESULTS

24-h mortality was 0.9%, 1.0%, 1.1%, 1.7%, 2.3%, 2.0% and 3.5% for patients with LTRS of 0,1,2,3,4, 5 and ≥ 6, respectively (P < 0.001). 30-day mortality was 3.5%, 4.2%, 4.9%, 6.2%, 7.6%, 7.2% and 10.1% respectively (P < 0.001). 1-year mortality was 8.6%, 10.8%, 12.9%, 13.9%, 18.5%, 20.3% and 28.6% respectively (P < 0.001). 10-year survival was 61%, 56%, 57%, 54%, 47%, and 31% for patients with 0, 1, 2, 3, 4, 5 and ≥ 6 points respectively (P < 0.001).

CONCLUSION

Perioperative mortality and long-term survival of patients undergoing LT can be accurately estimated at the time of listing by the LTRS.

Preoperative Stratification of Liver Transplant Recipients: Validation of the LTRS

BACKGROUND

The liver transplant risk score (LTRS) was developed to stratify 90-day mortality of patients referred for liver transplantation (LT). We aimed to validate the LTRS using a new cohort of patients.

METHODS

The LTRS stratifies the risk of 90-day mortality of LT recipients based on their age, body mass index, diabetes, model for end-stage liver disease (MELD) score, and need for dialysis. We assessed the performance of the LTRS using a new cohort of patients transplanted in the United States between July 2013 and June 2017. Exclusion criteria were age <18 years, ABO incompatibility, redo or multivisceral transplants, partial grafts, malignancies other than hepatocellular carcinoma and fulminant hepatitis.

RESULTS

We found a linear correlation between the number of points of the LTRS and 90-day mortality. Among 18 635 recipients, 90-day mortality was 2.7%, 3.8%, 5.2%, 4.8%, 6.7%, and 9.3% for recipients with 0, 1, 2, 3, 4, and ≥5 points (P < 0.001). The LTRS also stratified 1-year mortality that was 5.5%, 7.7%, 9.9%, 9.3%, 10.8%, and 15.4% for 0, 1, 2, 3, 4, and ≥5 points (P < 0.001). An inverse correlation was found between the LTRS and 4-year survival that was 82%, 79%, 78%, 82%, 78%, and 66% for patients with 0, 1, 2, 3, 4, and ≥5 points (P < 0.001). The LTRS remained an independent predictor after accounting for recipient sex, ethnicity, cause of liver disease, donor age, cold ischemia time, and waiting time.

CONCLUSIONS

The LTRS can stratify the short- and long-term outcomes of LT recipients at the time of their evaluations irrespective of their gender, ethnicity, and primary cause of liver disease.

Artificial neural network and bioavailability of the immunosuppression drug

PURPOSE OF REVIEW

The success of organ transplant is determined by number of demographic, clinical, immunological and genetic variables. Artificial intelligence tools, such as artificial neural networks (ANNs) or classification and regression trees (CART) can handle multiple independent variables and predict the dependent variables by deducing the complex nonlinear relationships between variables.

RECENT FINDINGS

In the last two decades, several researchers employed these tools to identify donor-recipient matching pairs, to optimize immunosuppressant doses, to predict allograft survival and to minimize adverse drug reactions. These models showed better performance characteristics than the empirical dosing strategies in terms of sensitivity, specificity, overall accuracy, or area under the curve of receiver-operating characteristic curves. The performance of the models was dependent directly on the input variables. Recent studies identified protein biomarkers and pharmacogenetic determinants of immunosuppressants as additional variables that increase the precision in prediction. Accessibility of medical records, proper follow-up of transplant cases, deep understanding of pharmacokinetic and pharmacodynamic pathways of immunosuppressant drugs coupled with genomic and proteomic markers are essential in developing an effective artificial intelligence platform for transplantation.

SUMMARY

Artificial intelligence has a greater clinical utility both in pretransplantation and posttransplantation periods to get favourable clinical outcomes, thus ensuring successful graft survival.

Artificial neural network models to predict nodal status in clinically node-negative breast cancer

BACKGROUND

Sentinel lymph node biopsy (SLNB) is standard staging procedure for nodal status in breast cancer, but lacks therapeutic benefit for patients with benign sentinel nodes. For patients with positive sentinel nodes, individualized surgical strategies are applied depending on the extent of nodal involvement. Preoperative prediction of nodal status is thus important for individualizing axillary surgery avoiding unnecessary surgery. We aimed to predict nodal status in clinically node-negative breast cancer and identify candidates for SLNB omission by including patient-related and pathological characteristics into artificial neural network (ANN) models.

METHODS

Patients with primary breast cancer were consecutively included between January 1, 2009 and December 31, 2012 in a prospectively maintained pathology database. Clinical- and radiological data were extracted from patient's files and only clinically node-negative patients constituted the final study cohort. ANN-based models for nodal prediction were constructed including 15 risk variables for nodal status. Area under the receiver operating characteristic curve (AUC) and Hosmer-Lemeshow goodness-of-fit test (HL) were used to assess performance and calibration of three predictive ANN-based models for no lymph node metastasis (N0), metastases in 1-3 lymph nodes (N1) and metastases in ≥ 4 lymph nodes (N2). Linear regression models for nodal prediction were calculated for comparison.

RESULTS

Eight hundred patients (N0, n = 514; N1, n = 232; N2, n = 54) were included. Internally validated AUCs for N0 versus N+ was 0.740 (95% CI = 0.723-0.758); median HL was 9.869 (P = 0.274), for N1 versus N0, 0.705 (95% CI = 0.686-0.724; median HL: 7.421; P = 0.492) and for N2 versus N0 and N1, 0.747 (95% CI = 0.728-0.765; median HL: 9.220; P = 0.324). Tumor size and vascular invasion were top-ranked predictors of all three end-points, followed by estrogen receptor status and lobular cancer for prediction of N2. For each end-point, ANN models showed better discriminatory performance than multivariable logistic regression models. Accepting a false negative rate (FNR) of 10% for predicting N0 by the ANN model, SLNB could have been abstained in 27.25% of patients with clinically node-negative axilla.

CONCLUSIONS

In this retrospective study, ANN showed promising result as decision-supporting tools for estimating nodal disease. If prospectively validated, patients least likely to have nodal metastasis could be spared SLNB using predictive models.

TRIAL REGISTRATION

Registered in the ISRCTN registry with study ID ISRCTN14341750 . Date of registration 23/11/2018. Retrospectively registered.

Validation of artificial neural networks as a methodology for donor-recipient matching for liver transplantation

In 2014, we reported a model for donor-recipient (D-R) matching in liver transplantation (LT) based on artificial neural networks (ANNs) from a Spanish multicenter study (Model for Allocation of Donor and Recipient in España [MADR-E]). The aim is to test the ANN-based methodology in a different European health care system in order to validate it. An ANN model was designed using a cohort of patients from King's College Hospital (KCH; n = 822). The ANN was trained and tested using KCH pairs for both 3- and 12-month survival models. End points were probability of graft survival (correct classification rate [CCR]) and nonsurvival (minimum sensitivity [MS]). The final model is a rule-based system for facilitating the decision about the most appropriate D-R matching. Models designed for KCH had excellent prediction capabilities for both 3 months (CCR-area under the curve [AUC] = 0.94; MS-AUC = 0.94) and 12 months (CCR-AUC = 0.78; MS-AUC = 0.82), almost 15% higher than the best obtained by other known scores such as Model for End-Stage Liver Disease and balance of risk. Moreover, these results improve the previously reported ones in the multicentric MADR-E database. In conclusion, the use of ANN for D-R matching in LT in other health care systems achieved excellent prediction capabilities supporting the validation of these tools. It should be considered as the most advanced, objective, and useful tool to date for the management of waiting lists. Liver Transplantation 24 192-203 2018 AASLD.

Use of artificial intelligence as an innovative donor-recipient matching model for liver transplantation: results from a multicenter Spanish study

BACKGROUND & AIMS

There is an increasing discrepancy between the number of potential liver graft recipients and the number of organs available. Organ allocation should follow the concept of benefit of survival, avoiding human-innate subjectivity. The aim of this study is to use artificial-neural-networks (ANNs) for donor-recipient (D-R) matching in liver transplantation (LT) and to compare its accuracy with validated scores (MELD, D-MELD, DRI, P-SOFT, SOFT, and BAR) of graft survival.

METHODS

64 donor and recipient variables from a set of 1003 LTs from a multicenter study including 11 Spanish centres were included. For each D-R pair, common statistics (simple and multiple regression models) and ANN formulae for two non-complementary probability-models of 3-month graft-survival and -loss were calculated: a positive-survival (NN-CCR) and a negative-loss (NN-MS) model. The NN models were obtained by using the Neural Net Evolutionary Programming (NNEP) algorithm. Additionally, receiver-operating-curves (ROC) were performed to validate ANNs against other scores.

RESULTS

Optimal results for NN-CCR and NN-MS models were obtained, with the best performance in predicting the probability of graft-survival (90.79%) and -loss (71.42%) for each D-R pair, significantly improving results from multiple regressions. ROC curves for 3-months graft-survival and -loss predictions were significantly more accurate for ANN than for other scores in both NN-CCR (AUROC-ANN=0.80 vs. -MELD=0.50; -D-MELD=0.54; -P-SOFT=0.54; -SOFT=0.55; -BAR=0.67 and -DRI=0.42) and NN-MS (AUROC-ANN=0.82 vs. -MELD=0.41; -D-MELD=0.47; -P-SOFT=0.43; -SOFT=0.57, -BAR=0.61 and -DRI=0.48).

CONCLUSIONS

ANNs may be considered a powerful decision-making technology for this dataset, optimizing the principles of justice, efficiency and equity. This may be a useful tool for predicting the 3-month outcome and a potential research area for future D-R matching models.

A hybrid model for predicting the prevalence of schistosomiasis in humans of Qianjiang City, China

Backgrounds/Objective

Schistosomiasis is still a major public health problem in China, despite the fact that the government has implemented a series of strategies to prevent and control the spread of the parasitic disease. Advanced warning and reliable forecasting can help policymakers to adjust and implement strategies more effectively, which will lead to the control and elimination of schistosomiasis. Our aim is to explore the application of a hybrid forecasting model to track the trends of the prevalence of schistosomiasis in humans, which provides a methodological basis for predicting and detecting schistosomiasis infection in endemic areas.

Methods

A hybrid approach combining the autoregressive integrated moving average (ARIMA) model and the nonlinear autoregressive neural network (NARNN) model to forecast the prevalence of schistosomiasis in the future four years. Forecasting performance was compared between the hybrid ARIMA-NARNN model, and the single ARIMA or the single NARNN model.

Results

The modelling mean square error (MSE), mean absolute error (MAE) and mean absolute percentage error (MAPE) of the ARIMA-NARNN model was 0.1869×10⁻⁴, 0.0029, 0.0419 with a corresponding testing error of 0.9375×10⁻⁴, 0.0081, 0.9064, respectively. These error values generated with the hybrid model were all lower than those obtained from the single ARIMA or NARNN model. The forecasting values were 0.75%, 0.80%, 0.76% and 0.77% in the future four years, which demonstrated a no-downward trend.

Conclusion

The hybrid model has high quality prediction accuracy in the prevalence of schistosomiasis, which provides a methodological basis for future schistosomiasis monitoring and control strategies in the study area. It is worth attempting to utilize the hybrid detection scheme in other schistosomiasis-endemic areas including other infectious diseases.

Assessing the effect of quantitative and qualitative predictors on gastric cancer individuals survival using hierarchical artificial neural network models

Background

There are numerous unanswered questions in the application of artificial neural network models for analysis of survival data. In most studies, independent variables have been studied as qualitative dichotomous variables, and results of using discrete and continuous quantitative, ordinal, or multinomial categorical predictive variables in these models are not well understood in comparison to conventional models.

Objectives

This study was designed and conducted to examine the application of these models in order to determine the survival of gastric cancer patients, in comparison to the Cox proportional hazards model.

Patients and Methods

We studied the postoperative survival of 330 gastric cancer patients who suffered surgery at a surgical unit of the Iran Cancer Institute over a five-year period. Covariates of age, gender, history of substance abuse, cancer site, type of pathology, presence of metastasis, stage, and number of complementary treatments were entered in the models, and survival probabilities were calculated at 6, 12, 18, 24, 36, 48, and 60 months using the Cox proportional hazards and neural network models. We estimated coefficients of the Cox model and the weights in the neural network (with 3, 5, and 7 nodes in the hidden layer) in the training group, and used them to derive predictions in the study group. Predictions with these two methods were compared with those of the Kaplan-Meier product limit estimator as the gold standard. Comparisons were performed with the Friedman and Kruskal-Wallis tests.

Results

Survival probabilities at different times were determined using the Cox proportional hazards and a neural network with three nodes in the hidden layer; the ratios of standard errors with these two methods to the Kaplan-Meier method were 1.1593 and 1.0071, respectively, revealed a significant difference between Cox and Kaplan-Meier (P < 0.05) and no significant difference between Cox and the neural network, and the neural network and the standard (Kaplan-Meier), as well as better accuracy for the neural network (with 3 nodes in the hidden layer). Probabilities of survival were calculated using three neural network models with 3, 5, and 7 nodes in the hidden layer, and it has been observed that none of the predictions was significantly different from results with the Kaplan-Meier method and they appeared more comparable towards the last months (fifth year). However, we observed better accuracy using the neural network with 5 nodes in the hidden layer. Using the Cox proportional hazards and a neural network with 3 nodes in the hidden layer, we found enhanced accuracy with the neural network model.

Conclusions

Neural networks can provide more accurate predictions for survival probabilities compared to the Cox proportional hazards mode, especially now that advances in computer sciences have eliminated limitations associated with complex computations. It is not recommended in order to adding too many hidden layer nodes because sample size related effects can reduce the accuracy. We recommend increasing the number of nodes to a point that increased accuracy continues (decrease in mean standard error), however increasing nodes should cease when a change in this trend is observed.

Enhancing the fever workup utilizing a multi-technique modeling approach to diagnose infections more accurately

BACKGROUND

Differentiation between infectious and non-infectious etiologies of the systemic inflammatory response syndrome (SIRS) in trauma patients remains elusive. We hypothesized that mathematical modeling in combination with computerized clinical decision support would assist with this differentiation. The purpose of this study was to determine the capability of various mathematical modeling techniques to predict infectious complications in critically ill trauma patients and compare the performance of these models with a standard fever workup practice (identifying infections on the basis of fever or leukocytosis).

METHODS

An 18-mo retrospective database was created using information collected daily from critically ill trauma patients admitted to an academic surgical and trauma intensive care unit. Two hundred forty-three non-infected patient-days were chosen randomly to combine with the 243 infected-days, which created a modeling sample of 486 patient-days. Utilizing ten variables known to be associated with infectious complications, decision trees, neural networks, and logistic regression analysis models were created to predict the presence of urinary tract infections (UTIs), bacteremia, and respiratory tract infections (RTIs). The data sample was split into a 70% training set and a 30% testing set. Models were compared by calculating sensitivity, specificity, positive predictive value, negative predictive value, overall accuracy, and discrimination.

RESULTS

Decision trees had the best modeling performance, with a sensitivity of 83%, an accuracy of 82%, and a discrimination of 0.91 for identifying infections. Both neural networks and decision trees outperformed logistic regression analysis. A second analysis was performed utilizing the same 243 infected days and only those non-infected patient-days associated with negative microbiologic cultures (n = 236). Decision trees again had the best modeling performance for infection identification, with a sensitivity of 79%, an accuracy of 83%, and a discrimination of 0.87.

CONCLUSION

The use of mathematical modeling techniques beyond logistic regression can improve the robustness and accuracy of predicting infections in critically ill trauma patients. Decision tree analysis appears to have the best potential to use in assisting physicians in differentiating infectious from non-infectious SIRS.

Artificial neural networks: useful aid in diagnosing acute appendicitis

BACKGROUND

[corrected] The purpose of the study was to assess the role of artificial neural networks (ANNs) in the diagnosis of appendicitis in patients presenting with acute right iliac fossa (RIF) pain and comparing its performance with the assessment made by experienced clinicians and the Alvarado score.

METHODS

After training and testing an ANN, data from 60 patients presenting with suspected appendicitis over a 6-month period to a teaching hospital was collected prospectively. Accuracy of diagnosing appendicitis by the clinician, the Alvarado score, and the ANN was compared.

RESULTS

The sensitivity, specificity, and positive and negative predictive values of the ANN were 100%, 97.2%, 96.0%, and 100% respectively. The ability of the ANN to exclude accurately the diagnosis of appendicitis in patients without true appendicitis was statistically significant compared to the clinical performance (p=0.031) and Alvarado score of >or=6 (p=0.004) and nearly significant compared to the Alvarado score of >or=7 (p=0.063).

CONCLUSIONS

ANNs can be an effective tool for accurately diagnosing appendicitis and may reduce unnecessary appendectomies.

Assessment of gastric cancer survival: using an artificial hierarchical neural network

This study is designed to assess the application of neural networks in comparison to the Kaplan-Meier and Cox proportional hazards model in the survival analysis. Three hundred thirty gastric cancer patients admitted to and surgically treated were assessed and their post-surgical survival was determined. The observed baseline survival was determined with the three methods of Kaplan-Meier product limit estimator, Cox and the neural network and results were compared. Then the binary independent variables were entered into the model. Data were randomly divided into two groups of 165 each to test the models and assess the reproducibility. The Chi-square test and the multiple logistic model were used to ensure the groups were similar and the data was divided randomly. To compare subgroups, we used the log-rank test. In the next step, the probability of survival in different periods was computed based on the training group data using the Cox proportional hazards and a neural network and estimating Cox coefficient values and neural network weights (with 3 nodes in hidden layer). Results were used for predictions in the test group data and these predictions were compared using the Kaplan-Meier product limit estimator as the gold standard. Friedman and Kruskal-Wallis tests were used for comparisons as well. All statistical analyses were performed using SPSS version 11.5, Matlab version 7.2, Statistica version 6.0 and S_PLUS 2000. The significance level was considered 5% (alpha = 0.05). The three methods used showed no significance difference in base survival probabilities. Overall, there was no significant difference among the survival probabilities or the trend of changes in survival probabilities calculated with the three methods, but the 4 year (48th month) and 4.5 year (54th month) survival rates were significantly different with Cox compared to standard and estimated probabilities in the neural network (p < 0.05). Kaplan-Meier and Cox showed almost similar results for the baseline survival probabilities, but results with the neural network were different: higher probabilities up to the 4th year, then comparable with the other two methods. Estimates from Cox proportional hazards and the neural network with three nodes in hidden layer were compared with the estimate from the Kaplan-Meier estimator as the gold standard. Neither comparison showed statistically significant differences. The standard error ratio of the two estimate groups by Cox and the neural network to Kaplan-Meier were no significant differences, it indicated that the neural network was more accurate. Although we do not suggest neural network methods to estimate the baseline survival probability, it seems these models is more accurately estimated as compared with the Cox proportional hazards, especially with today's advanced computer sciences that allow complex calculations. These methods are preferable because they lack the limitations of conventional models and obviate the need for unnecessary assumptions including those related to the proportionality of hazards and linearity.

Predicting Kidney Transplant Survival Using Tree-Based Modeling

Predicting the outcome of kidney transplantation is clinically important and computationally challenging. The goal of this project was to develop the models predicting probability of kidney allograft survival at 1, 3, 5, 7, and 10 years. Kidney transplant data from the United States Renal Data System (January 1, 1990, to December 31, 1999, with the follow-up through December 31, 2000) were used (n = 92,844). Independent variables included recipient demographic and anthropometric data, end-stage renal disease course, comorbidity information, donor data, and transplant procedure variables. Tree-based models predicting the probability of the allograft survival were generated using roughly two-thirds of the data (training set), with the remaining one-third left aside to be used for models validation (testing set). The prediction of the probability of graft survival in the independent testing dataset achieved a good correlation with the observed survival (r = 0.94, r = 0.98, r = 0.99, r = 0.93, and r = 0.98) and relatively high areas under the receiving operator characteristic curve (0.63, 0.64, 0.71, 0.82, and 0.90) for 1-, 3-, 5-, 7-, and 10-year survival prediction, respectively. The models predicting the probability of 1-, 3-, 5-, 7-, and 10-year allograft survival have been validated on the independent dataset and demonstrated performance that may suggest implementation in clinical decision support system.

Design of an Artificial Neural Network for Diagnosis of Facial Pain Syndromes

A classification scheme for facial pain syndromes describing seven categories has previously been proposed. Based on this classification scheme and a binomial (yes/no) facial pain questionnaire, we have designed and trained an artificial neural network (ANN) and as an initial feasibility assessment of such an ANN system examined its ability to recognize and correctly diagnose patients with different facial pain syndromes. One hundred patients with facial pain were asked to respond to a facial pain questionnaire at the time of their initial visit. After interview, an independent diagnosis was assigned to each patient. The patients' responses to the questionnaire and their diagnoses were input to an ANN. The ANN was able to retrospectively predict the correct diagnosis for 95 of 100 patients (95%), and prospectively determine a correct diagnosis of trigeminal neuralgia Type 1 with 84% sensitivity and 83% specificity in 43 new patients. The ability of the ANN to accurately predict a correct diagnosis for the remaining types of facial pain was limited by our clinic sample size and hence less exposure to those categories. This is the first demonstration of the utilization of an ANN to diagnose facial pain syndromes.

Prediction of Cyclosporine Blood Levels in Heart Transplantation Patients Using a Pharmacokinetic Model Identified by Evolutionary Algorithms

BACKGROUND

Artificial intelligence (AI)-based computation methods have been recently shown to be applicable in several clinical diagnostic fields. The purpose of this study was to introduce a novel AI method called evolutionary algorithms (EAs) to clinical predictions. The technique was used to create a pharmacokinetic model for the prediction of whole blood levels of cyclosporine (CyA).

METHODS

One hundred one adult cardiac transplant recipients were randomly selected and included in this study. All patients had been receiving oral cyclosporine twice daily, and the trough levels in whole blood were measured by monoclonal-specific radioimmunoassay. An evolutionary algorithm (EA)-based software tool was trained with pre- and post-operative variables from 64 patients. The results of this process were then tested on data sets from 37 patients.

RESULTS

The mean value of the predicted CyA level throughout the measurement period for the test data was 175 +/- 27 ng/ml, which compared well with the mean observed CyA level of 180 +/- 31 ng/ml. The system bias expressed as the mean percent error (MPE) for the training and test data sets were 7.1 +/- 5.4% (0.1% to 26.7%) and 8.0 +/- 6.7% (0.8% to 28.8%), respectively. The prediction accuracy ranged from 80% to 90%. The correlation coefficient between predicted and observed CyA concentration for the training data were 0.93 (p < 0.001) and for the test data were 0.85 (p < 0.001), respectively.

CONCLUSIONS

The results of this study suggest that the use of evolutionary algorithms to identify pharmacokinetic models yields accurate prediction of cyclosporine whole blood levels in heart transplant recipients. This and other similar technologies should be considered as future clinical tools to reduce costs in our health systems.

Artificial neural networks as prediction tools in the critically ill

The past 25 years have witnessed the development of improved tools with which to predict short-term and long-term outcomes after critical illness. The general paradigm for constructing the best known tools has been the logistic regression model. Recently, a variety of alternative tools, such as artificial neural networks, have been proposed, with claims of improved performance over more traditional models in particular settings. However, these newer methods have yet to demonstrate their practicality and usefulness within the context of predicting outcomes in the critically ill.

Self-Organizing Maps Can Determine Outcome and Match Recipients and Donors at Orthotopic Liver Transplantation

BACKGROUND

There is a relative lack of donor organs for liver transplantation. Ideally, to maximize the utility of those livers that are offered, donor and recipient characteristics should be matched to ensure the best possible posttransplant survival of the recipient.

METHODS

With prospectively collected data on 827 patients receiving a primary liver graft for chronic liver disease, we used a self-organizing map (SOM) (one form of a neural network) to predict outcome after transplantation using both donor and recipient factors. The SOM was then validated using a data set of 2622 patients undergoing transplantation in the United Kingdom at other centers.

RESULTS

SOM analysis using 72 inputs and two survival intervals (3 and 12 months) yielded three neurons with either higher or lower probabilities of survival. The model was validated using the independent data set. With 20 patients on the waiting list and 10 sequential donor livers, it was possible to demonstrate that the model could be used to identify which potential recipients were likely to benefit most from each liver offered.

CONCLUSIONS

With this approach to matching donor livers and recipients, it is possible to inform transplant clinicians about the optimum use of donor livers and thereby effectively make the best use of a scarce resource.

Expanded criteria donor grafts for deceased donor liver transplantation under the MELD system: a decision analysis

Expanded criteria donor (ECD) liver grafts have a higher likelihood of primary graft failure (PGF) compared with standard criteria donor (SCD) grafts. Given a choice between an available ECD graft versus waiting for an SCD graft that may not always become available, what should liver transplant candidates do? The study's aim was to estimate 1-year survival comparing immediate ECD liver grafting with waiting for an SCD organ. Using UNOS data, published literature estimates, and expert opinion, we constructed a Markov decision analytic model to estimate survival while waiting for an SCD transplant and survival with immediate ECD transplant. Sensitivity analyses were performed by varying model parameters individually and simultaneously with a second-order Monte Carlo simulation. For all patients with MELD scores >20, survival was higher with immediate ECD transplant despite the additional increased risk for PGF. Survival was better with an immediate ECD transplant unless the probability of PGF exceeded 23%, 72%, and 88% for recipients with MELD scores of 11-20, 21-25, and 26-30 respectively. For patients with MELD scores >30, the survival benefit with the immediate ECD strategy persisted at even higher rates of PGF. In conclusion, our results suggest that, despite the higher risk for PGF, transplantation with an available ECD graft should be preferred over waiting for an SCD organ for patients with advanced MELD scores. At less advanced MELD scores, the survival benefit depends on the risk of PGF associated with the ECD organ.

A model to predict survival at one month, one year, and five years after liver transplantation based on pretransplant clinical characteristics

Reliable models that could predict outcome of liver transplantation (LT) may guide physicians to advise their patients of immediate and late survival chances and may help them to optimize organ use. The objective of this study was to develop user-friendly models to predict short and long-term mortality after LT in adults based on pre-LT recipient characteristics. The United Network for Organ Sharing (UNOS) transplant registry (n = 38,876) from 1987 to 2001 was used to develop and validate the model. Two thirds of patients were randomized to develop the model (the modeling group), and the remaining third was randomized to cross-validate (the cross-validation group) it. Three separate models, using multivariate logistic regression analysis, were created and validated to predict survival at 1 month, 1 year, and 5 years. Using the total severity scores of patients in the modeling group, a predictive model then was created, and the predicted probability of death as a function of total score then was compared in the cross-validation group. The independent variables that were found to be very significant for 1 month and 1 year survival were age, body mass index (BMI), UNOS status 1, etiology, serum bilirubin (for 1 month and 1 year only), creatinine, and race (only for 5 years). The actual deaths in the cross-validation group followed very closely the predicted survival graph. The chi-squared goodness-of-fit test confirmed that the model could predict mortality reliably at 1 month, 1 year, and 5 years. We have developed and validated user-friendly models that could reliably predict short-term and long-term survival after LT.

Pretransplant model to predict posttransplant survival in liver transplant patients

OBJECTIVE

To develop a prognostic model that determines patient survival outcomes after orthotopic liver transplantation (OLT) using readily available pretransplant variables.

SUMMARY BACKGROUND DATA

The current liver organ allocation system strongly favors organ distribution to critically ill recipients who exhibit poor survival outcomes following OLT. A severely limited organ resource, increasing waiting list deaths, and rising numbers of critically ill recipients mandate an organ allocation system that balances disease severity with survival outcomes. Such goals can be realized only through the development of prognostic models that predict survival following OLT.

METHODS

Variables that may affect patient survival following OLT were analyzed in hepatitis C (HCV) recipients at the authors' center, since HCV is the most common indication for OLT. The resulting patient survival model was examined and refined in HCV and non-HCV patients in the United Network for Organ Sharing (UNOS) database. Kaplan-Meier methods, univariate comparisons, and multivariate Cox proportional hazard regression were employed for analyses.

RESULTS

Variables identified by multivariate analysis as independent predictors for patient survival following primary transplantation of adult HCV recipients in the last 10 years at the authors' center were entered into a prognostic survival model to predict patient survival. Accordingly, mortality was predicted by 0.0293 (recipient age) + 1.085 (log10 recipient creatinine) + 0.289 (donor female gender) + 0.675 urgent UNOS - 1.612 (log10 recipient creatinine times urgent UNOS). The above variables, in addition to donor age, total bilirubin, prothrombin time (PT), retransplantation, and warm and cold ischemia times, were applied to the UNOS database. Of the 46,942 patients transplanted over the last 10 years, 25,772 patients had complete data sets. An eight-factor model that accurately predicted survival was derived. Accordingly, the mortality index posttransplantation = 0.0084 donor age + 0.019 recipient age + 0.816 log creatinine + 0.0044 warm ischemia (in minutes) + 0.659 (if second transplant) + 0.10 log bilirubin + 0.0087 PT + 0.01 cold ischemia (in hours). Thus, this model is applicable to first or second liver transplants. Patient survival rates based on model-predicted risk scores for death and observed posttransplant survival rates were similar. Additionally, the model accurately predicted survival outcomes for HCV and non-HCV patients.

CONCLUSIONS

Posttransplant patient survival can be accurately predicted based on eight straightforward factors. The balanced application of a model for liver transplant survival estimate, in addition to disease severity, as estimated by the model for end-stage liver disease, would markedly improve survival outcomes and maximize patients' benefits following OLT.

Pretransplant model to predict posttransplant survival in liver transplant patients

OBJECTIVE

To develop a prognostic model that determines patient survival outcomes after orthotopic liver transplantation (OLT) using readily available pretransplant variables.

SUMMARY BACKGROUND DATA

The current liver organ allocation system strongly favors organ distribution to critically ill recipients who exhibit poor survival outcomes following OLT. A severely limited organ resource, increasing waiting list deaths, and rising numbers of critically ill recipients mandate an organ allocation system that balances disease severity with survival outcomes. Such goals can be realized only through the development of prognostic models that predict survival following OLT.

METHODS

Variables that may affect patient survival following OLT were analyzed in hepatitis C (HCV) recipients at the authors' center, since HCV is the most common indication for OLT. The resulting patient survival model was examined and refined in HCV and non-HCV patients in the United Network for Organ Sharing (UNOS) database. Kaplan-Meier methods, univariate comparisons, and multivariate Cox proportional hazard regression were employed for analyses.

RESULTS

Variables identified by multivariate analysis as independent predictors for patient survival following primary transplantation of adult HCV recipients in the last 10 years at the authors' center were entered into a prognostic survival model to predict patient survival. Accordingly, mortality was predicted by 0.0293 (recipient age) + 1.085 (log10 recipient creatinine) + 0.289 (donor female gender) + 0.675 urgent UNOS - 1.612 (log10 recipient creatinine times urgent UNOS). The above variables, in addition to donor age, total bilirubin, prothrombin time (PT), retransplantation, and warm and cold ischemia times, were applied to the UNOS database. Of the 46,942 patients transplanted over the last 10 years, 25,772 patients had complete data sets. An eight-factor model that accurately predicted survival was derived. Accordingly, the mortality index posttransplantation = 0.0084 donor age + 0.019 recipient age + 0.816 log creatinine + 0.0044 warm ischemia (in minutes) + 0.659 (if second transplant) + 0.10 log bilirubin + 0.0087 PT + 0.01 cold ischemia (in hours). Thus, this model is applicable to first or second liver transplants. Patient survival rates based on model-predicted risk scores for death and observed posttransplant survival rates were similar. Additionally, the model accurately predicted survival outcomes for HCV and non-HCV patients.

CONCLUSIONS

Posttransplant patient survival can be accurately predicted based on eight straightforward factors. The balanced application of a model for liver transplant survival estimate, in addition to disease severity, as estimated by the model for end-stage liver disease, would markedly improve survival outcomes and maximize patients' benefits following OLT.

Development of a model for prediction of survival in pediatric trauma patients: comparison of artificial neural networks and logistic regression

BACKGROUND/PURPOSE

There is a paucity of outcome prediction models for injured children. Using the National Pediatric Trauma Registry (NPTR), the authors developed an artificial neural network (ANN) to predict pediatric trauma death and compared it with logistic regression (LR).

METHODS

Patients in the NPTR from 1996 through 1999 were included. Models were generated using LR and ANN. A data search engine was used to generate the ANN with the best fit for the data. Input variables included anatomic and physiologic characteristics. There was a single output variable: probability of death. Assessment of the models was for both discrimination (ROC area under the curve) and calibration (Lemeshow-Hosmer C-Statistic).

RESULTS

There were 35,385 patients. The average age was 8.1 +/- 5.1 years, and there were 1,047 deaths (3.0%). Both modeling systems gave excellent discrimination (ROC A(z): LR = 0.964, ANN = 0.961). However, LR had only fair calibration, whereas the ANN model had excellent calibration (L/H C stat: LR = 36, ANN = 10.5).

CONCLUSIONS

The authors were able to develop an ANN model for the prediction of pediatric trauma death, which yielded excellent discrimination and calibration exceeding that of logistic regression. This model can be used by trauma centers to benchmark their performance in treating the pediatric trauma population.

Modeling medical prognosis: survival analysis techniques

Medical prognosis has played an increasing role in health care. Reliable prognostic models that are based on survival analysis techniques have been recently applied to a variety of domains, with varying degrees of success. In this article, we review some methods commonly used to model time-oriented data, such as Kaplan-Meier curves, Cox proportional hazards, and logistic regression, and discuss their applications in medical prognosis. Nonlinear, nonparametric models such as neural networks have increasingly been used for building prognostic models. We review their use in several medical domains and discuss different implementation strategies. Advantages and disadvantages of these methods are outlined, as well as pointers to pertinent literature.

Preoperative factors associated with outcome and their impact on resource use in 1148 consecutive primary liver transplants

BACKGROUND

Hepatic transplantation is a highly effective but costly treatment for end-stage hepatic dysfunction. One approach to improve efficiency in the use of scarce organs for transplantation is to identify preoperative factors that are associated with poor outcome posttransplantation. This may assist both in selecting patients optimal for transplantation and in identifying strategies to improve survival.

METHODS

In the present work, we retrospectively reviewed consecutive liver transplants performed at the University of California at Los Angeles during a 6-year period and determined preoperative variables that were associated with outcome in primary grafts. In addition, we used the hospital's cost accounting database to determine the impact of these variables on the degree of resource use by high-risk patients.

RESULTS

We found five variables to have independent prognostic value in predicting graft survival after primary liver transplantation: (1) donor age, (2) recipient age, (3) donor sodium, (4) recipient creatinine, and (5) recipient ventilator requirement pretransplant. Recipient ventilator requirement and elevated creatinine were associated with significant increases in resource use during the transplant admission.

CONCLUSIONS

Patients at high risk for graft failure and costly transplants can be identified preoperatively by a set of parameters that are readily available, noninvasive, and inexpensive. Selection of recipients on the basis of these data would improve the efficiency of liver transplantation and reduce its cost.

Validity of an artificial neural network in predicting discharge destination from a postacute geriatric rehabilitation unit

OBJECTIVE

To develop an artificial neural network (ANN) designed to predict discharge destination from postacute geriatric rehabilitation units.

DESIGN

Nonconcurrent prospective study.

SETTING

Postacute geriatric rehabilitation units: a 20-bed unit in a nonproprietary skilled nursing facility and a 40-bed unit in a suburban private facility.

PATIENTS

Consecutive sample of 661 patients admitted between January 1995 and February 1999, including a derivation group of 452 patients and a validation group of 209 patients.

INTERVENTIONS

A feed-forward, back-propagation neural network to predict discharge destination.

MAIN OUTCOME MEASURE

Discharge destination from postacute geriatric rehabilitation.

RESULTS

An ANN was trained on clinical pattern set derived from 452 patients and validated prospectively on 209 consecutive patients admitted to postacute geriatric rehabilitation units. The neural network achieved a sensitivity of 85.7% (95% confidence interval [CI], 83.7-89.4) and specificity of 94.1% (95% CI, 84.4-99.1) in identifying discharge destination with a corresponding area under the curve of 95.7% (95% CI, 92.1-98.3).

CONCLUSION

An ANN can predict discharge to the community postacute rehabilitation with a high degree of accuracy. It could have particular value to predict return to the community for older adults with multiple comorbidities after an acute hospitalization.

An artificial neural network as a model for prediction of survival in trauma patients: validation for a regional trauma area

BACKGROUND

To develop and validate an artificial neural network (ANN) for predicting survival of trauma patients based on standard prehospital variables, emergency room admission variables, and Injury Severity Score (ISS) using data derived from a regional area trauma system, and to compare this model with known trauma scoring systems.

PATIENT POPULATION

The study was composed of 10,609 patients admitted to 24 hospitals comprising a seven-county suburban/rural trauma region adjacent to a major metropolitan area. The data was generated as part of the New York State trauma registry. Study period was from January 1993 through December 1996 (1993-1994: 5,168 patients; 1995: 2,768 patients; 1996: 2,673 patients).

METHODS

A standard feed-forward back-propagation neural network was developed using Glasgow Coma Scale, systolic blood pressure, heart rate, respiratory rate, temperature, hematocrit, age, sex, intubation status, ICD-9-CM Injury E-code, and ISS as input variables. The network had a single layer of hidden nodes. Initial network development of the model was performed on the 1993-1994 data. Subsequent models were generated using the 1993, 1994, and 1995 data. The model was tested first on the 1995 and then on the 1996 data. The ANN model was tested against Trauma and Injury Severity Score (TRISS) and ISS using the receiver operator characteristic (ROC) area under the curve [ROC-A(z)], Lemeshow-Hosmer C-statistic, and calibration curves.

RESULTS

The ANN showed good clustering of the data, with good separation of nonsurvivors and survivors. The ROCA(z) was 0.912 for the ANN, 0.895 for TRISS, and 0.766 for ISS. The ANN exceeded TRISS with respect to calibration (Lemeshow-Hosmer C-statistic: 7.4 for ANN; 17.1 for TRISS). The prediction of survivors was good for both models. The ANN exceeded TRISS in nonsurvivor prediction.

CONCLUSION

An ANN developed for trauma patients using prehospital, emergency room admission data, and ISS gave good prediction of survival. It was accurate and had excellent calibration. This study expands our previous results developed at a single Level I trauma center and shows that an ANN model for predicting trauma deaths can be applied across hospitals with good results

Neuro-computing versus linear statistical techniques applied to liver transplant monitoring: a comparative study

This paper explores the potential for the application of neurocomputing in on-line monitoring in the liver transplantation domain. It extends our previously documented work to provide both an assessment of the performance gains achievable by incorporating temporal and dynamical information about the measurements made on a patient as well as presenting a novel computerized clinical decision aid for this domain. A comparison of the performance of linear and nonlinear classification system is made and used to motivate the final selection of the diagnostic inputs.

Major complications after angioplasty in patients with chronic renal failure: a comparison of predictive models

Novel modeling approaches were investigated to predict major complications in patients with chronic renal failure (CRF) or end-stage renal disease (ESRD) undergoing percutaneous transluminal coronary angioplasty (PTCA). The following hypotheses were explored: (1) Pre-angioplasty patient risk factors, demographic characteristics and procedural information may be used to predict major complications after PTCA; and (2) Rough sets and artificial neural nets (ANN) may be used to build models that are better than standard logistic regression models. Several variables were found to be predictive of major complications for patients with CRF or ESRD undergoing PTCA. The presence of shock at presentation portends poor outcome but congestive heart failure and prior history of myocardial infarction increases the risk tenfold and 25-fold, respectively. The discriminatory ability of the ANN model was better than both Rough Sets and Logistic Regression for the test set.

Short-term and long-term outcome prediction with the Acute Physiology and Chronic Health Evaluation II system after orthotopic liver transplantation

OBJECTIVE

To evaluate the relationship between the postoperative Acute Physiology and Chronic Health Evaluation (APACHE) II score and mortality at hospital discharge and at 1 yr in liver transplant recipients.

POPULATION

Adult orthotopic liver transplant (OLTX) recipients (n = 599) admitted to the intensive care unit postoperatively at a university hospital.

METHODS

The cohort was split randomly into development and validation sets. Three models were compared for each end point: a) the original APACHE II slope with the original APACHE II postgastrointestinal surgery intercept; b) the original APACHE II slope with an OLTX-specific intercept generated from the development set; and c) an OLTX-specific slope and intercept generated from the development set. Goodness-of-fit and calibration were assessed by the Hosmer-Lemeshow C statistic (where p>.05 suggests good fit) and standardized mortality ratios. Discrimination was assessed by receiver operator characteristic area under the curve analysis.

MEASUREMENTS AND MAIN RESULTS

Hospital and 1-yr mortality rates were 9.9% and 15.9%, respectively. The APACHE II score was strongly associated with mortality (chi-square, p<.0001), but when used with the original equation, it significantly overestimated hospital mortality (standardized mortality ratio, 0.73 [confidence interval, 0.58-0.99]). Using the OLTX-specific approaches, goodness-of-fit for both hospital and 1-yr mortality was good (p = .2-.57) but discrimination was only moderate (receiver operator characteristic area under the curve, 0.675-0.723).

CONCLUSIONS

APACHE II is a good predictor of short- and long-term mortality after liver transplantation, especially when using OLTX-specific coefficients. Because fit and calibration were better than discrimination, APACHE II will be most useful in the prediction of risk for groups of patients (e.g., in clinical trials or institutional comparisons) rather than for individuals. This study raises the possibility that APACHE II may be useful for long-term mortality prediction in other critically ill populations. The overestimation of mortality using the original equation suggests that orthotopic liver transplantation, by reversing the underlying pathophysiology, may modify risk.

Predicting active pulmonary tuberculosis using an artificial neural network

BACKGROUND

Nosocomial outbreaks of tuberculosis (TB) have been attributed to unrecognized pulmonary TB. Accurate assessment in identifying index cases of active TB is essential in preventing transmission of the disease.

OBJECTIVES

To develop an artificial neural network using clinical and radiographic information to predict active pulmonary TB at the time of presentation at a health-care facility that is superior to physicians' opinion.

DESIGN

Nonconcurrent prospective study.

SETTING

University-affiliated hospital.

PARTICIPANTS

A derivation group of 563 isolation episodes and a validation group of 119 isolation episodes.

INTERVENTIONS

A general regression neural network (GRNN) was used to develop the predictive model.

MEASUREMENTS

Predictive accuracy of the neural network compared with clinicians' assessment.

RESULTS

Predictive accuracy was assessed by the c-index, which is equivalent to the area under the receiver operating characteristic curve. The GRNN significantly outperformed the physicians' prediction, with calculated c-indices (+/- SEM) of 0.947 +/- 0.028 and 0.61 +/- 0.045, respectively (p < 0.001). When the GRNN was applied to the validation group, the corresponding c-indices were 0. 923 +/- 0.056 and 0.716 +/- 0.095, respectively.

CONCLUSION

An artificial neural network can identify patients with active pulmonary TB more accurately than physicians' clinical assessment.

Costs and outcomes of liver transplantation in adults: a prospective, 1-year, follow-up study. GRETHECO study group

BACKGROUND

Orthotopic liver transplantation (OLT) is widely used to treat patients with end-stage liver disease. However, data on the cost of the procedure are fragmentary. We evaluated the costs, as calculated from resource use, and outcomes of OLT in adults, from registration on the transplant waiting list to the end of the 1st-year of follow-up after the transplant.

METHODS

Two parallel cohort studies were conducted from 1994 to 95. All patients ages 18 years and older, on the waiting list (n=33) according to national criteria or having undergone transplants (n=38) were followed for 1 year or until either the transplant (waiting list cohort) or death (waiting list and transplantation cohorts).

RESULTS

Eighty percent of the patients undergoing transplants were alive after 1 year, and no patient died while on the waiting list. However, the estimated cost of the procedure was high: more than 55,000 pound silver for the 1st year after OLT, to be added to 5,500 pound silver for evaluation and further costs motivated by the planned transplant during an average 6.5 months on the waiting list. Age over 40 and a baseline Child-Pugh score of 10 and over were predictive of high costs. The proportion of costs associated with immunosuppressive therapy and rejection were very high.

CONCLUSIONS

This medical and economic cohort study suggests that OLT is still expensive; the study identifies sources of extra cost that could be limited either by improved selection of patients or, in the future, by technological advances in immunosuppressive therapy that help avoid medical complications. It also suggests the situation is precarious, with outcomes and costs being very sensitive to variation in graft availability.

Neural network applications in physical medicine and rehabilitation

The purpose of this article is to provide an overview of neural networks and their applications in physical medicine and rehabilitation. Conventional statistical models may present certain limitations that can be overcome by neural networks. We show what neural networks are, how they "learn" regularities from the data, and how they can classify previously unseen cases. We present advantages and disadvantages of using neural networks and compare them with regression models. We explain how neural networks can be used as statistical tools for making inferences using the example of a prognostic model that predicts ambulation after spinal cord injury.

Predicting cytomegalovirus disease after renal transplantation: an artificial neural network approach

Outcome prediction is becoming increasingly important in medicine, but when a resource is scarce the need for accurate prediction becomes acute. Prediction based on biostatistical models has been in use for many years, but in areas such as renal transplantation their results have been disappointing. Recently however, there has been growing interest in the use of artificial neural networks for prediction. The creation of a large database containing high quality data on renal transplantation patients in Wales offers an ideal opportunity to research a new area viz., the ability of these techniques to accurately predict outcomes such as the appearance of disease in transplant recipients or the time to graft failure. This paper describes the use of neural networks to identify patients who risk the development of cytomegalovirus disease--a significant cause of mortality and morbidity in these patients. The neural networks we examined produced overall correct classifications well in excess of 80% in each of the two groups involved, diseased and non-diseased. These predictions are a considerable improvement on current methods and encourage the belief that renal transplantation data may respond well to analysis by neural networks.

Understanding artificial neural networks and exploring their potential applications for the practicing urologist

Artificial neural networks (ANNs) are complex mathematical models that are distantly based on the human neuronal structure. They are capable of modeling elaborate biologic systems without making assumptions based on statistical distributions. Preliminary work has been reported on their application in urology. The initial results have been promising, particularly as an additional tool in the detection of early prostate cancer using the ProstAsure Index, which has been the most extensively studied urologic ANN to date. We review the basic concepts behind ANNs and examine currently existing and potential future applications of this new dynamic technology both in urology and in general clinical medicine.

MHC haplotype analysis by artificial neural networks

Conventional matching is based on numbers of alleles shared between donor and recipient. This approach, however, ignores the degree of relationship between alleles and haplotypes, and therefore the actual degree of difference. To address this problem, we have compared family members using a block matching technique which reflects differences in genomic sequences. All parents and siblings had been genotyped using conventional MHC typing so that haplotypes could be assigned and relatives could be classified as sharing 0, 1 or 2 haplotypes. We trained an Artificial Neural Network (ANN) with subjects from 6 families (85 comparisons) to distinguish between relatives. Using the outputs of the ANN, we developed a score, the Histocompatibility Index (HI), as a measure of the degree of difference. Subjects from a further 3 families (106 profile comparisons) were tested. The HI score for each comparison was plotted. We show that the HI score is trimodal allowing the definition of three populations corresponding to approximately 0, 1 or 2 haplotype sharing. The means and standard deviations of the three populations were found. As expected, comparisons between family members sharing 2 haplotypes resulted in high HI scores with one exception. More interestingly, this approach distinguishes between the 1 and 0 haplotype groups, with some informative exceptions. This distinction was considered too difficult to attempt visually. The approach provides promise in the quantification of degrees of histocompatibility.

Artificial neural networks applied to outcome prediction for colorectal cancer patients in separate institutions

BACKGROUND

Artificial neural networks are computer programs that can be used to discover complex relations within data sets. They permit the recognition of patterns in complex biological data sets that cannot be detected with conventional linear statistical analysis. One such complex problem is the prediction of outcome for individual patients treated for colorectal cancer. Predictions of outcome in such patients have traditionally been based on population statistics. However, these predictions have little meaning for the individual patient. We report the training of neural networks to predict outcome for individual patients from one institution and their predictive performance on data from a different institution in another region.

METHODS

5-year follow-up data from 334 patients treated for colorectal cancer were used to train and validate six neural networks designed for the prediction of death within 9, 12, 15, 18, 21, and 24 months. The previously trained 12-month neural network was then applied to 2-year follow-up data from patients from a second institution; outcome was concealed. No further training of the neural network was undertaken. The network's predictions were compared with those of two consultant colorectal surgeons supplied with the same data.

FINDINGS

All six neural networks were able to achieve overall accuracy greater than 80% for the prediction of death for individual patients at institution 1 within 9, 12, 15, 18, 21, and 24 months. The mean sensitivity and specificity were 60% and 88%. When the neural network trained to predict death within 12 months was applied to data from the second institution, overall accuracy of 90% (95% CI 84-96) was achieved, compared with the overall accuracy of the colorectal surgeons of 79% (71-87) and 75% (66-84).

INTERPRETATION

The neural networks were able to predict outcome for individual patients with colorectal cancer much more accurately than the currently available clinicopathological methods. Once trained on data from one institution, the neural networks were able to predict outcome for patients from an unrelated institution.

Sequential versus standard neural networks for pattern recognition: an example using the domain of coronary heart disease

The goal of this study was to compare standard and sequential neural network models for recognition of patterns of disease progression. Medical researchers who perform prognostic modeling usually oversimplify the problem by choosing a single point in time to predict outcomes (e.g. death in 5 years). This approach not only fails to differentiate patterns of disease progression, but also wastes important information that is usually available in time-oriented research data bases. The adequate use of sequential neural networks can improve the performance of prognostic systems if the interdependencies among prognoses at different intervals of time are explicitly modeled. In such models, predictions for a certain interval of time (e.g. death within 1 year) are influenced by predictions made for other intervals, and prognostic survival curves that provide consistent estimates for several points in time can be produced. We developed a system of neural network models that makes use of time-oriented data to predict development of coronary heart disease (CHD), using a set of 2594 patients. The output of the neural network system was a prognostic curve representing survival without CHD, and the inputs were the values of demographic, clinical, and laboratory variables. The system of neural networks was trained by backpropagation and its results were evaluated in test sets of previously unseen cases. We showed that, by explicitly modeling time in the neural network architecture, the performance of the prognostic index, measured by the area under the receiver operating characteristic (ROC) curve, was significantly improved (p < 0.05).

Cardiac risk stratification in renal transplantation using a form of artificial intelligence

The purpose of this study was to determine if an expert network, a form of artificial intelligence, could effectively stratify cardiac risk in candidates for renal transplant. Input into the expert network consisted of clinical risk factors and thallium-201 stress test data. Clinical risk factor screening alone identified 95 of 189 patients as high risk. These 95 patients underwent thallium-201 stress testing, and 53 had either reversible or fixed defects. The other 42 patients were classified as low risk. This algorithm made up the "expert system," and during the 4-year follow-up period had a sensitivity of 82%, specificity of 77%, and accuracy of 78%. An artificial neural network was added to the expert system, creating an expert network. Input into the neural network consisted of both clinical variables and thallium-201 stress test data. There were 5 hidden nodes and the output (end point) was cardiac death. The expert network increased the specificity of the expert system alone from 77% to 90% (p < 0.001), the accuracy from 78% to 89% (p < 0.005), and maintained the overall sensitivity at 88%. An expert network based on clinical risk factor screening and thallium-201 stress testing had an accuracy of 89% in predicting the 4-year cardiac mortality among 189 renal transplant candidates.

A comparison of Cox proportional hazards and artificial neural network models for medical prognosis

Modeling survival of populations and establishing prognoses for individual patients are important activities in the practice of medicine. For patients with diseases that may extend for several years, in particular, accurate assessment of survival probabilities is essential. New methods, such as neural networks, have been used increasingly to model disease progression. Their advantages and disadvantages, when compared to statistical methods such as Cox proportional hazards, have seldom been explored in real-world data. In this study, we compare the performances of a Cox model and a neural network model that are used as prognostic tools for a set of people living with AIDS. We modeled disease progressions for patients who had AIDS (according to the 1993 CDC definition) in a set of 588 patients in California, using data from the ATHOS project. We divided the study population into 10 training and 10 test sets and evaluated the prognostic accuracy of a Cox proportional hazards model and of a neural network model by determining sensitivities, specificities, positive and negative predictive values for an arbitrary threshold (0.5), and the areas under the receiver operating characteristics (ROC) curves that utilized all possible thresholds for intervals of 1 yr following the diagnosis of AIDS. There was no evidence that the Cox model performed better than did the neural network model or vice versa, but the former method had the advantage of providing some insight on which variables were most influential for prognosis. Nevertheless, it is likely that the assumptions required by the Cox model may not be satisfied in all data sets, justifying the use of neural networks in certain cases.

Hepatic Retransplantation--an analysis of risk factors associated with outcome

Hepatic retransplantation is controversial because the results are inferior to primary transplants and organs are so scarce. To determine the factors that are associated with poor outcome within the first year following retransplantation, we performed a multivariate analysis, using stepwise logistic regression, of 418 hepatic retransplantations performed at a single institution from November 1987 to December 1993. The minimum follow-up was 1 year. Seven variables were found to be independently associated with subsequent graft failure (defined as either patient death or retransplantation): donor age (odds ratio 2.2 for each 10-year increase over age 45, 95% CI 1.3 to 3.7), female donor sex (odds ratio 1.7, 95% CI 1.05 to 2.7), recipient age (odds ratio 1.6 for each 10-year increase over age 45,95% CI 1.2 to 2.8), need for preoperative mechanical ventilation (odds ratio 1.8, 95% CI 1.1 to 2.9), pretransplant serum creatinine (odds ratio 1.24 for each increase of 1 mg/dl, 95% CI 1.1 to 1.4), pretransplant total serum bilirubin (odds ratio 1.4 for each 10-mg/dl increase over 15 mg/dl, 95% CI 1.1 to 1.8), and the primary immunosuppressant, using tacrolimus as the reference category (odds ratio for cyclosporine-based immunosuppression 3.9, 95% CI 2.3 to 6.8). Although not part of the logistic regression model, the timing of retransplantation was also found to be important, with the overall probability of failure increasing from 0.58 on day 0 to a peak of 0.8 on day 38 and decreasing slowly after that. The implications of these results regarding the appropriateness of retransplantation are discussed.