BACKGROUND Given its narrow therapeutic range, digoxin's pharmacokinetic parameters in infants are difficult to predict due to variation in birth weight and gestational age, especially for critically ill newborns. There is limited evidence to support the safety and dosage requirements of digoxin, let alone to predict its concentrations in infants. This study aimed to compare the concentrations of digoxin predicted by traditional regression modeling and artificial neural network (ANN) modeling for newborn infants given digoxin for clinically significant patent ductus arteriosus (PDA). METHODS A retrospective chart review was conducted to obtain data on digoxin use for clinically significant PDA in a neonatal intensive care unit. Newborn infants who were given digoxin and had digoxin concentration(s) within the acceptable range were identified as subjects in the training model and validation datasets, accordingly. Their demographics, disease, and medication information, which were potentially associated with heart failure, were used for model training and analysis of digoxin concentration prediction. The models were generated using backward standard multivariable linear regressions (MLRs) and a standard backpropagation algorithm of ANN, respectively. The common goodness-of-fit estimates, receiver operating characteristic curves, and classification of sensitivity and specificity of the toxic concentrations in the validation dataset obtained from MLR or ANN models were compared to identify the final better predictive model. RESULTS Given the weakness of correlations between actual observed digoxin concentrations and pre-specified variables in newborn infants, the performance of all ANN models was better than that of MLR models for digoxin concentration prediction. In particular, the nine-parameter ANN model has better forecasting accuracy and differentiation ability for toxic concentrations. CONCLUSION The nine-parameter ANN model is the best alternative than the other models to predict serum digoxin concentrations whenever therapeutic drug monitoring is not available. Further cross-validations using diverse samples from different hospitals for newborn infants are needed.

中文翻译：

---

通过人工神经网络预测新生儿重症监护病房中婴儿的血清地高辛浓度。

背景技术由于地高辛的治疗范围狭窄，由于出生体重和胎龄的变化，尤其是重症新生儿，很难预测婴儿的地高辛的药代动力学参数。仅有有限的证据支持地高辛的安全性和剂量要求，更不用说预测婴儿中地高辛的浓度了。这项研究旨在比较通过传统回归模型和人工神经网络（ANN）模型预测的地高辛对具有临床意义的动脉导管未闭（PDA）的新生儿的地高辛浓度。方法进行回顾性图表审查，以获取有关在新生儿重症监护室中具有临床意义的PDA使用地高辛的数据。因此，在训练模型和验证数据集中将接受了地高辛且地高辛浓度在可接受范围内的新生婴儿识别为受试者。他们的人口统计学，疾病和药物治疗信息可能与心力衰竭有关，被用于模型训练和地高辛浓度预测分析。分别使用后向标准多元线性回归（MLR）和ANN的标准反向传播算法生成模型。比较了从MLR或ANN模型获得的验证数据集中的通用拟合优度估计值，接收器工作特性曲线以及毒性浓度的敏感性和特异性的分类，以识别出最终的更好的预测模型。结果由于新生儿中实际观察到的地高辛浓度与预先设定的变量之间的相关性较弱，因此在预测地高辛浓度时，所有ANN模型的性能均优于MLR模型。尤其是，九参数人工神经网络模型对毒物浓度具有更好的预测准确性和区分能力。结论九参数人工神经网络模型是其他模型中预测地高辛浓度的最佳替代方法，只要无法进行治疗药物监测即可。需要使用来自不同医院的不同样本对新生儿进行进一步的交叉验证。尤其是，九参数人工神经网络模型对毒物浓度具有更好的预测准确性和区分能力。结论九参数人工神经网络模型是其他模型中预测地高辛浓度的最佳替代方法，只要无法进行治疗药物监测即可。需要使用来自不同医院的不同样本对新生儿进行进一步的交叉验证。尤其是，九参数人工神经网络模型对毒物浓度具有更好的预测准确性和区分能力。结论九参数人工神经网络模型是其他模型中预测地高辛浓度的最佳替代方法，只要无法进行治疗药物监测即可。需要使用来自不同医院的不同样本对新生儿进行进一步的交叉验证。