This study was performed in order to characterize once more the linear and non-linear relationship between several clinical and biological variables and the response to epoetin beta and to test the performance of non-linear mathematical models, based on artificial neural networks, in the prediction of the erythropoietin dose required for an individual patient. Our results show that ANNs which could be implemented in wards are clearly superior to linear regressions or the nephrologist's opinion in predicting the erythropoietin dose for an individual patient.
The first observation of this study is that linear regressions are less performant than ANNs as predictive tools. Indeed, as illustrated in Figure 3, in contrary to non-linear mathematical models based on ANNs, the best performing linear regression did not demonstrate any significant ability to predict erythropoietin responsiveness. This performance gap confirms a non-linear relationship between factors influencing both erythropoietin responsiveness and need in individual patients. As highlighted in previous studies,[41,42] on the basis of the results of the linear regressions, the administration route of erythropoietin significantly influenced the erythropoietin dose: patients treated with intravenous erythropoietin required a higher erythropoietin dose (16.0% more) and had lower haemoglobin (0.53 g/dL less).
Analysing the AIMSEOC data pool, the main variables influencing the mean erythropoietin dose needed to obtain a haemoglobin of 11.5 g/dL were weight, drug administration route (subcutaneous vs. intravenous), age and ferritin (see Figure 2 Panel A). Among these variables the relevance of the correlation with the weight has to be considered cautiously since the epoetin dose in the AIMSEOC database was indexed to body weight and in a previous study the absence of justification for body weight adjusted dosage was demonstrated. Confirming previous epidemiological and experimental data, the analysis of the EOC subgroup allowed the identification of further variables showing a relevant non linear relationship: particular attention should be given to pH, Kt/V, PTH and CRP (see Figure 2 Panel B).[9,10,13,14,15]
The good performance of the ANN built using as input variables weight, epoetin administration route, age, presence or absence of cardiomyopathy and creatinine is shown in Figure 3 (ROC curves). For a specificity of 50%, the sensitivity of ANNs compared with linear regressions in predicting the erythropoietin dose to reach the haemoglobin target was 78 vs. 44% (P < 0.001). Considering that the cited variables are included in the blood tests usually performed before starting an epoetin substitution, additional information about erythropoietin responsiveness could be easy obtained without supplementary costs.
Curiously enough not one of the cited variables contributed significantly in the building of the model structured to predict the monthly adaptations in epoetin dose in individual patients. This means that, as suggested by pharmacodynamic models for other drugs and as confirmed by the results of previous studies,[24,26] the monthly fluctuations in haemoglobin as a function of the erythropoietin dose over a 3-month period are indirectly expressing all the other tested parameters related, in an analogous non linear mathematical model, to erythropoietin responsiveness. Of note, a large intra- and inter-individual variability in the requirements of erythropoietin (17.5 ± 19.2 and 35.3 ± 34.0 % respectively), to be referred at least in part to the inclusion in the study even of patients with intercurrent illnesses susceptible to influence the haemoglobin value, making the prediction of the ideal dose particularly difficult, was found in our database.
Compared with the nephrologists in charge of the patients, following the European best practice guidelines, the best performing ANN built to predict the monthly adaptations in epoetin dose on the basis of the haemoglobin and of the epoetin prescribed in the previous two months would allow the detection of 48 vs. 25% of the patients treated with an insufficient dose with a specificity of 92 vs. 83% (positive and negative predictive values 71 vs. 39 and 82 vs. 73% respectively). Only in 2 cases (0.8% of the tested group) the follow-up haemoglobin of the patients selected to be treated with a higher epoetin dose would have been > 12.0 g/dL (12.1 and 12.8 g/dL respectively) without adaptation in the dose. This finding, compared with the performance of the nephrologists in the same group (the epoetin dose would have been increased in 4 patients with a follow-up haemoglobin > 12.0 g/dL without adaptation in the dose), offers sufficient guarantees for the application of the selected ANNs in the clinical setting. Furthermore, compared with previous studies ( Table 4 ), the present one was conducted on a larger multicentric group of patients (432 from 29 dialysis units) with a strong inter-individual variability. This fact should guarantee the applicability of the models beyond the studied population. However, considering that increasing the prediction power of the nephrologists is not the only condition needed to promote efficiency in erythropoietin prescription, the real impact of our computer assisted tool has to be evaluated in a prospective randomized trial.
Coming back to the choice of using ANNs as a non linear adaptive learning machine for individualizing epoetin dosage regimens, their usefulness by clinical ward oriented nephrologists has been demonstrated once again. However even if compared to other computer assisted mathematical models for non linear adaptive modeling ANNs are easy to use and to access and tolerate both missing data and errors in individual variables well, their user friendliness contrasts with the still persisting difficulties in correctly evaluating the reliability of the obtained functions.[34,35,36]
The next step will be to include in the electronic documentation of the dialysis patients in use in our centres individualized models automatically warning the nephrologists about the need and modality of adaptations in the epoetin dose.
BMC Nephrology © 2006 Gabutti et al; licensee BioMed Central Ltd.
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Cite this: Would Artificial Neural Networks Implemented in Clinical Wards Help Nephrologists in Predicting Epoetin Responsiveness? - Medscape - Sep 01, 2006.