- Email: [email protected]
Determining tacrolimus blood concentrations in the intensive care unit
Abstracts
P-10 Determining tacrolimus blood concentrations in the intensive care unit Jelle Van Eyck a, J. Ramon a, F Guiza a, G. Meyfroidt b, M. Bruynooghe a, G. Van den Berghe b a Department of Computer Science, K.U. Leuven, Celestijnenlaan 200A, B-3001 Heverlee, Belgium b Department of Intensive Care Medicine, K.U. Leuven. University Hospital Gasthuisberg, Herestraat 49, B-3000 Leuven, Belgium
Objectives: Tacrolimus is an immunosuppressive drug that is used after organ transplant to reduce the activity of the patient's immune system and so lower the risk of organ rejection. Its defining properties are its narrow therapeutic range, possible toxic adverse effects, and a blood concentration with a wide intraindividual and interindividual variability. To ensure clinical efficacy and avoid toxic adverse effects, it is of pivotal importance that the tacrolimus blood concentration is monitored rigidly. In our intensive care unit, tacrolimus blood concentrations are measured on a daily basis (8 AM), and tacrolimus medication is administered twice a day (8 AM and 8 PM). In this study, we use machine learning techniques to determine the tacrolimus blood concentration of a patient after his evening dose of tacrolimus. Accurate predictions of the tacrolimus blood concentration allow us to optimize the dosing regime, resulting in optimal clinical effect. Methods: A total of 270 potentially clinically relevant parameters (including admission data, parameters indicating renal dysfunction, parameters indicating liver dysfunction, other physiological parameters, administered medication, etc) were selected from 188 patients who underwent organ transplantation. To acquire a general model, we limited the number of examples per patient to 5, resulting in a total number of 756 examples to learn from. Relevance vector machines (RVMs) were used to determine which of these clinical parameters are relevant for predicting future tacrolimus blood concentrations and how they interact. One of the difficulties is the selection of clinically relevant parameters. However, because RVMs are known to produce sparse results, parameter selection can be guided by the resulting models. Furthermore, for the model to be easily interpretable, we chose to use RVMs in input space only. Results: Given the relatively small number of examples contained in the data set, we used 10-fold cross-validation to evaluate the different models. This was done by calculating the mean absolute error between the predicted values and the corresponding target values. In clinical practice, a combination of the previous tacrolimus blood concentration and the 2 most recent tacrolimus doses are used to predict the future tacrolimus blood concentration of a patient. As a baseline, we created an RVM using exactly these clinical parameters. This resulted in a mean absolute error of 2.30 ng/mL. Extending the previous model and including the 6 most recent doses of tacrolimus instead of only the 2 most recent ones resulted in an improvement of the mean absolute error to 2.24 ng/ mL. Adding other clinical parameters did not improve the results any further. Except for the case of Diflucan (Pfizer, New York, NY), which, when coadministered with tacrolimus, resulted in a significant rise of the tacrolimus blood concentration. In this case, the mean absolute error was found to be 2.22 ng/mL. Conclusions: In this study, we identified a number of clinical parameters that are clearly predictive for tacrolimus blood concentrations. These parameters are the previous tacrolimus
e11 blood concentrations, the 6 most recent doses of tacrolimus, and the administration of Diflucan. In future work, we will further extend our techniques to improve our results even further. doi: 10.1016/j.jcrc.2012.01.032
P-11 Generalized multifractal analysis of heart rate variability recordings with a large number of arrhythmia Jan Gieraltowski a, Jan Zebrowski a, Rafal Baranowski b a Faculty of Physics, Warsaw University of Technology, Warszawa, Poland b The Cardinal Stefan Wyszyński Institute of Cardiology, Warszawa, Poland
Objectives: The regulation of human heart rate is the result of many inputs, for example, the activity of the sympathetic and parasympathetic nervous system, respiration and its control, or such pathologies as ectopic activity or delayed conduction of cardiac tissue—each having its own characteristic time scale and magnitude. The MultiFractal Detrended Fluctuation Analysis (MF-DFA) method used by us allows to assess the effect of the different controls systems and pathologies. Because it requires stationarity, the method is applied in the literature to heart rate variability recordings with less than 5% of arrhythmia, and we wanted to improve this number. Methods: We analyzed the published MF-DFA method using synthetic data and chosen R-R intervals series. We developed an original, generalized version of the MF-DFA method—multiscale multifractal analysis. We found that the calculation of the f(α) curve is a major source of artifacts. We, thus, focused on the dependence of the local Hurst exponent h on the multifractal parameter q: h(q), and we allowed it to depend on the scale s. In the standard MFDFA, the time scale s is fixed, somewhat arbitrarily (usually from 50 intervals up to 500). Thus, we obtained the h(q,s) dependence— a surface—the shape of which tells us what is the magnitude of the fluctuations the R-R intervals have in different time scales (different frequency bands). Results: Multiscale multifractal analysis was found to be immune to noise contamination of the data (we tested up to 50% of noise). It also allows to study heart rate variability with an arbitrary level of arrhythmia required for clinical applications. We analyzed fifty-one 24-hour recordings of heart rate variability (36 males aged 16-64 years, 15 females aged 11-57 years: 42 healthy persons, 9 cardiac arrest cases including 5 without organic heart disease). Also, preliminary results for a group of subjects with aortic valve stenosis just before surgery and after follow-up will be presented. We did not remove arrhythmia from the recordings. We limited the study to the night hours to avoid arbitrary daytime activity. Conclusions: Our mathematical criterion was able to distinguish, in a blind test, healthy subjects from the high-risk cardiac arrest cases including those without organic disease. The different peculiarities of each recording have a unique effect on the results of the multiscale MF-DFA analysis; for example, the occurrence of arrhythmia may readily be identified from the results. Thus, the new method allows to recognize and assign a complexity measure to features of the heart rate variability, which hitherto went unnoticed when using standard, linear diagnostic methods and MF-DFA. doi: 10.1016/j.jcrc.2012.01.033
P-10 Determining tacrolimus blood concentrations in the intensive care unit Jelle Van Eyck a, J. Ramon a, F Guiza a, G. Meyfroidt b, M. Bruynooghe a, G. Van den Berghe b a Department of Computer Science, K.U. Leuven, Celestijnenlaan 200A, B-3001 Heverlee, Belgium b Department of Intensive Care Medicine, K.U. Leuven. University Hospital Gasthuisberg, Herestraat 49, B-3000 Leuven, Belgium
Objectives: Tacrolimus is an immunosuppressive drug that is used after organ transplant to reduce the activity of the patient's immune system and so lower the risk of organ rejection. Its defining properties are its narrow therapeutic range, possible toxic adverse effects, and a blood concentration with a wide intraindividual and interindividual variability. To ensure clinical efficacy and avoid toxic adverse effects, it is of pivotal importance that the tacrolimus blood concentration is monitored rigidly. In our intensive care unit, tacrolimus blood concentrations are measured on a daily basis (8 AM), and tacrolimus medication is administered twice a day (8 AM and 8 PM). In this study, we use machine learning techniques to determine the tacrolimus blood concentration of a patient after his evening dose of tacrolimus. Accurate predictions of the tacrolimus blood concentration allow us to optimize the dosing regime, resulting in optimal clinical effect. Methods: A total of 270 potentially clinically relevant parameters (including admission data, parameters indicating renal dysfunction, parameters indicating liver dysfunction, other physiological parameters, administered medication, etc) were selected from 188 patients who underwent organ transplantation. To acquire a general model, we limited the number of examples per patient to 5, resulting in a total number of 756 examples to learn from. Relevance vector machines (RVMs) were used to determine which of these clinical parameters are relevant for predicting future tacrolimus blood concentrations and how they interact. One of the difficulties is the selection of clinically relevant parameters. However, because RVMs are known to produce sparse results, parameter selection can be guided by the resulting models. Furthermore, for the model to be easily interpretable, we chose to use RVMs in input space only. Results: Given the relatively small number of examples contained in the data set, we used 10-fold cross-validation to evaluate the different models. This was done by calculating the mean absolute error between the predicted values and the corresponding target values. In clinical practice, a combination of the previous tacrolimus blood concentration and the 2 most recent tacrolimus doses are used to predict the future tacrolimus blood concentration of a patient. As a baseline, we created an RVM using exactly these clinical parameters. This resulted in a mean absolute error of 2.30 ng/mL. Extending the previous model and including the 6 most recent doses of tacrolimus instead of only the 2 most recent ones resulted in an improvement of the mean absolute error to 2.24 ng/ mL. Adding other clinical parameters did not improve the results any further. Except for the case of Diflucan (Pfizer, New York, NY), which, when coadministered with tacrolimus, resulted in a significant rise of the tacrolimus blood concentration. In this case, the mean absolute error was found to be 2.22 ng/mL. Conclusions: In this study, we identified a number of clinical parameters that are clearly predictive for tacrolimus blood concentrations. These parameters are the previous tacrolimus
e11 blood concentrations, the 6 most recent doses of tacrolimus, and the administration of Diflucan. In future work, we will further extend our techniques to improve our results even further. doi: 10.1016/j.jcrc.2012.01.032
P-11 Generalized multifractal analysis of heart rate variability recordings with a large number of arrhythmia Jan Gieraltowski a, Jan Zebrowski a, Rafal Baranowski b a Faculty of Physics, Warsaw University of Technology, Warszawa, Poland b The Cardinal Stefan Wyszyński Institute of Cardiology, Warszawa, Poland
Objectives: The regulation of human heart rate is the result of many inputs, for example, the activity of the sympathetic and parasympathetic nervous system, respiration and its control, or such pathologies as ectopic activity or delayed conduction of cardiac tissue—each having its own characteristic time scale and magnitude. The MultiFractal Detrended Fluctuation Analysis (MF-DFA) method used by us allows to assess the effect of the different controls systems and pathologies. Because it requires stationarity, the method is applied in the literature to heart rate variability recordings with less than 5% of arrhythmia, and we wanted to improve this number. Methods: We analyzed the published MF-DFA method using synthetic data and chosen R-R intervals series. We developed an original, generalized version of the MF-DFA method—multiscale multifractal analysis. We found that the calculation of the f(α) curve is a major source of artifacts. We, thus, focused on the dependence of the local Hurst exponent h on the multifractal parameter q: h(q), and we allowed it to depend on the scale s. In the standard MFDFA, the time scale s is fixed, somewhat arbitrarily (usually from 50 intervals up to 500). Thus, we obtained the h(q,s) dependence— a surface—the shape of which tells us what is the magnitude of the fluctuations the R-R intervals have in different time scales (different frequency bands). Results: Multiscale multifractal analysis was found to be immune to noise contamination of the data (we tested up to 50% of noise). It also allows to study heart rate variability with an arbitrary level of arrhythmia required for clinical applications. We analyzed fifty-one 24-hour recordings of heart rate variability (36 males aged 16-64 years, 15 females aged 11-57 years: 42 healthy persons, 9 cardiac arrest cases including 5 without organic heart disease). Also, preliminary results for a group of subjects with aortic valve stenosis just before surgery and after follow-up will be presented. We did not remove arrhythmia from the recordings. We limited the study to the night hours to avoid arbitrary daytime activity. Conclusions: Our mathematical criterion was able to distinguish, in a blind test, healthy subjects from the high-risk cardiac arrest cases including those without organic disease. The different peculiarities of each recording have a unique effect on the results of the multiscale MF-DFA analysis; for example, the occurrence of arrhythmia may readily be identified from the results. Thus, the new method allows to recognize and assign a complexity measure to features of the heart rate variability, which hitherto went unnoticed when using standard, linear diagnostic methods and MF-DFA. doi: 10.1016/j.jcrc.2012.01.033