Evaluation of two cortical fraction estimation algorithms for the calculation of dynamic magnetic resonance renograms.

With the high resolution of dynamic magnetic resonance imaging (MRI) scans it is possible to measure cortical renograms directly, but due to partial volume effects this is impossible for medullary renograms. With weighted subtraction of the cortical renogram from a mixed renogram it becomes possible to extract the medullary renogram. For this subtraction the fraction of cortical tissue, present in the region of interest in which the mixed renogram is determined, has to be calculated. We have evaluated two algorithms for calculation of the cortical fraction. Both algorithms use the fact that during an interval after the start of the cortical enhancement no medullary enhancement occurs. One algorithm calculates the ratio between the slopes of both enhancement curves. The other is based on minimising the medullary signal values using a least squares error (LSE) method. Using a computer model of the renograms and measurements on real patients we analysed the accuracy of both methods and determined the best parameters for each.

Design and implementation of a framework to support the development of clinical guidelines.

This paper describes and discusses a framework that facilitates the development of clinical guideline application tasks. The framework, named GASTON covers all stages in the guideline development process, ranging from the definition of models that represent guidelines to the implementation of run-time systems that provide decision support, based on the guidelines that were developed during the earlier stages. The GASTON framework consists of (1) a newly developed guideline representation formalism that uses the concepts of primitives, problem-solving methods (PSMs) and ontologies to represent the guidelines of various complexity and granularity and different application domains, (2) a guideline authoring environment that enables guideline authors to define the guidelines, based on the newly developed guideline representation formalism and (3) a guideline execution environment that translates defined guidelines into a more efficient symbol level representation, which can be read in and processed by an execution time engine. The paper describes a number of design criteria that were formulated regarding the aspects of guideline representation, guideline authoring and guideline execution and explains the framework by example in terms of the four stages that were identified in the guideline development process and the tools that were developed to support each stage. It also shows examples of systems that were developed by means of the GASTON framework.

Movement correction of the kidney in dynamic MRI scans using FFT phase difference movement detection.

To measure cortical and medullary MR renograms, regions of interest (ROIs) are placed on the kidney in images acquired using dynamic MRI. Since native kidneys move with breathing, and breath-holding techniques are not feasible, movement correction is necessary. In this contribution we compare three correction methods, based on image matching, phase difference movement detection (PDMD), and cross-correlation, respectively. The PDMD-based method showed the best performance and was able to determine kidney movement in our test series in 68% of the scans with no visible deviation, and in 88% of the scans if a one-pixel deviation is considered acceptable.

Computer control versus manual control of systemic hypertension during cardiac surgery.

BACKGROUND: We recently demonstrated the feasibility of computer controlled infusion of vasoactive drugs for the control of systemic hypertension during cardiac surgery. The objective of the current study was to investigate the effects of computer controlled blood pressures on hemodynamic stability when compared to conventional manual control. METHOD: Systemic artery blood pressures were managed either by computer (80 patients) or by a well-trained anesthesiologist (80 patients). The vasodilator drugs sodium nitroprusside and nitroglycerin were used. Hemodynamic stability was determined from the standard deviation of the mean arterial pressure samples and from the percentages of time that arterial pressure was hypertensive or hypotensive. RESULTS: The average standard deviation of the mean arterial pressure samples was smaller for the computer controlled than for the manually controlled group: 7.5+/-2.2 (mean+/-SD) versus 8.9+/-2.3 mmHg (P<0.0001). The systemic artery pressure was less hypertensive and less hypotensive in the computer controlled than in the manually controlled group: 9.4+/-5.7 versus 13.1+/-6.0% (P<0.0001) and 8.0+/-5.9 versus 11.8+/-7.4% (P<0.0001), respectively. CONCLUSION: We conclude that, compared with manual control, computer control of systemic hypertension significantly improved hemodynamic stability during cardiac surgery.

The application of ontologies and problem-solving methods for the development of shareable guidelines.

Recently, studies have shown the benefits of using clinical guidelines in the practice of medicine. Computer-based clinical guidelines are increasingly applied in diverse areas such as policy development, utilization management, education, conduct of clinical trials, and workflow facilitation. This paper discusses some of the representations suggested in literature, discusses their weak and strong points, and demonstrates and discusses a new approach that extends earlier developed formalisms by combining primitives, ontologies and the use of problem-solving methods (PSMs). The approach is supported by a framework that facilitates the entire guideline authoring process. The paper demonstrates this framework and presents examples of guidelines, PSMs and systems that were developed by means of this approach. The overall goal of this approach is to improve the acceptance of shareable guidelines and decision support systems in daily care by facilitating the guideline acquisition and execution phases.

Reduction of noise in medullary renograms from dynamic MR images.

Dynamic magnetic resonance images of the kidney can be used to acquire separate renograms of the cortex and medulla. A high-quality cortical renogram can be determined directly from a region of interest (ROI) placed in the cortex. Due to partial volume effects, part of the signal from a ROI placed in the medulla is caused by cortical tissue. By subtracting a fraction of the cortical signal from the cortico-medullary signal, a purer medullary renogram can be obtained. A side effect of this subtraction is an increase in noise level. The noise level increases with larger partial volume fractions. Using a matched image filter, it is possible to exclude those areas from the ROI that have a high partial volume content, thus reducing the amount of cortical signal that has to be separated from the medullary signal. Noise reductions of up to 50% have been achieved in the medullary renogram, with an average reduction of 23%.

GASTON: an architecture for the acquisition and execution of clinical guideline-application tasks.

Recently, studies have shown the benefits of using clinical guidelines in the practice of medicine. There have been numerous efforts to develop clinical decision support systems that support guideline-based care in an automated fashion, covering a wide range of clinical settings and tasks. Despite these efforts, only a few systems progressed beyond the prototype stage and the research laboratory. For guideline-based clinical decision support systems to be successful, a balance must be made between intuitive but imprecise representations usually encountered by most of today's systems and representations that support a strong underlying clinical performance model. The project described in this paper tries to achieve such a balance. It presents the GASTON architecture that contains a set of reusable software components for the application of guidelines, including design-time components to facilitate the guideline authoring process based on guideline representation models along with execution-time components for building decision support systems that incorporate these guidelines. This architecture was used to develop several guideline representation models such as a rule-based representation to model rule-based guidelines and guideline representation models that address more complex tasks. Also, decision support systems that incorporate these models were developed with the architecture. For the representation and application of various classes of guidelines, rules were also viewed as instances of more complex tasks. By identifying similar characteristics of sets of rules, we developed several tasks such as a drug intera ction and drug contraindication task. Based on these models, we have developed and validated guidelines and decision support systems for use in several application domains such as intensive care, family physicians and psychiatry. In order to be able to represent more complex time-oriented plans, new guideline representation models are being developed.

A strategy for developing practice guidelines for the ICU using automated knowledge acquisition techniques.

OBJECTIVES: To implement practice guideline entry tools in a reminder system in order to provide decision support to health care workers in clinical care and emergency care environments. To design a knowledge acquisition environment that enables physicians to formulate, update, and verify guidelines without the assistance of a knowledge engineer. METHODS: We developed a knowledge acquisition environment for the Intensive Care Unit (ICU) consisting of 1) a graphical knowledge acquisition tool, 2) tools that perform logical and semantic tests on proposed guidelines, 3) a Patient Data Management System (PDMS) containing clinical patient data, and 4) an expert system that reminds ICU health care workers of inconsistencies between a treatment plan and implemented guidelines. Physicians enter the guidelines using the knowledge acquisition tool, after which consistency and correctness tests are performed on the guidelines. The guidelines are then transferred to the knowledge base of the reminder system and validated by applying the new guidelines to a large stored data set of previous patients. If the new guidelines are approved, they are exported to the reminder system that is used in daily practice. RESULTS: ICU physicians used the knowledge acquisition tool to enter 58 guidelines into the reminder system's knowledge base. These guidelines were tested on a data set consisting of 803 previously admitted patients. As a result, 27 guidelines fired at least once, generating 406 reminders in total. Of the 406 generated reminders, 356 (88%) were issued correctly and 50 (12%) were false alarms. The reminders that were issued correctly involved 3 situations: 1) the database contained inconsistent or incomplete information, 2) the actions or decisions of the health care workers were not the most appropriate ones, and 3) there was a potential risk involved. All false alarms were caused by the fact that the corresponding guidelines were not specific enough to handle certain exceptions. As a result of this analysis, the guidelines could be improved in such a way as to eliminate all false alarms. CONCLUSIONS: These first results demonstrate that this bottom-up knowledge acquisition strategy, implemented by the automated knowledge acquisition tools, enables medical specialists to improve the quality of computer support in an ICU without assistance of a knowledge engineer.

Automated infusion of vasoactive and inotropic drugs to control arterial and pulmonary pressures during cardiac surgery.

OBJECTIVE: To evaluate the feasibility of a closed-loop system for simultaneous control of systemic arterial and pulmonary artery blood pressures during cardiac surgery. DESIGN: Feasibility study. SETTING: The cardiac surgery operating room. PATIENTS: The performance of the multiple-drug closed-loop system was evaluated during cardiac surgery in 30 patients who required treatment with more than one vasoactive or inotropic drug. INTERVENTIONS: A multiple-drug closed-loop system integrated five single-drug blood pressure controllers. Arterial hypertension was controlled using sodium nitroprusside or nitroglycerin, arterial hypotension was controlled using noradrenaline or dobutamine, and pulmonary hypertension was controlled using nitroglycerin. The anesthesiologist selected target pressures and single-drug blood pressure controllers. The multiple-drug closed-loop system had a set of priority rules that automatically activated from the selected single-drug controllers the optimum single-drug controller for each hemodynamic state. Drug infusion rates of the nonactive controllers were kept constant. The initial knowledge that was used to construct the priority rules was obtained from standard anesthetic protocols on perioperative management of cardiac surgical patients. A supervisory computer program defined the actions to be taken in cases of infusion pump problems, invalid pressure measurements, and during unexpected increases and decreases in systemic arterial pressure. MEASUREMENTS AND MAIN RESULTS: The activation of single-drug controllers by the priority rules was accurate and fast. On average, a different single-drug controller was activated once every 7.2 mins. As a measure of variability, the average deviation of mean arterial pressure and mean pulmonary artery pressure from their target values was evaluated and was 8.6+/-4.0 and 4.4+/-4.0 mm Hg, respectively, before cardiopulmonary bypass and 8.0+/-3.6 and 2.4+/-0.9 mm Hg, respectively, after cardiopulmonary bypass. None of the single-drug controllers showed any signs of unstable response. CONCLUSION: Closed-loop control of both arterial and pulmonary pressures using multiple drugs is feasible during cardiac surgery.

Detection of dicrotic notch in arterial pressure signals.

OBJECTIVE: A novel algorithm to detect the dicrotic notch in arterial pressure signals is proposed. Its performance is evaluated using both aortic and radial artery pressure signals, and its robustness to variations in design parameters is investigated. METHODS: Most previously published dicrotic notch detection algorithms scan the arterial pressure waveform for the characteristic pressure change that is associated with the dicrotic notch. Aortic valves, however, are closed by the backwards motion of aortic blood volume. We developed an algorithm that uses arterial flow to detect the dicrotic notch in arterial pressure waveforms. Arterial flow is calculated from arterial pressure using simulation results with a three-element windkessel model. Aortic valve closure is detected after the systolic upstroke and at the minimum of the first negative dip in the calculated flow signal. RESULTS: In 7 dogs ejection times were derived from a calculated aortic flow signal and from simultaneously measured aortic flow probe data. A total of 86 beats was analyzed; the difference in ejection times was -0.6 +/- 5.4 ms (means +/- SD). The algorithm was further evaluated using 6 second epochs of radial artery pressure data measured in 50 patients. Model simulations were carried out using both a linear windkessel model and a pressure and age dependent nonlinear windkessel model. Visual inspection by an experienced clinician confirmed that the algorithm correctly identified the dicrotic notch in 98% (49 of 50) of the patients using the linear model, and 96% (48 of 50) of the patients using the nonlinear model. The position of the dicrotic notch appeared to be less sensitive to variations in algorithm's design parameters when a nonlinear windkessel model was used. CONCLUSIONS: The detection of the dicrotic notch in arterial pressure signals is facilitated by first calculating the arterial flow waveform from arterial pressure and a model of arterial afterload. The method is robust and reduces the problem of detecting a dubious point in a decreasing pressure signal to the detection of a well-defined minimum in a derived signal.

Clinical evaluation of an automatic blood pressure controller during cardiac surgery.

OBJECTIVE: During surgery, computers can be of great use to support the anesthesiologist in providing task automation. In this paper we describe a closed loop blood pressure controller and show the results of its clinical evaluation. METHODS: The controller is based on a simple and robust Proportional-Integral controller and a supervising, rule based, expert system. Adaptive control is necessary because the sensitivity of the patients to sodium nitroprusside varies over a wide range. Thirty-three clinical tests during cardiac surgery, including the cardiopulmonary bypass phase, were performed. RESULTS: On average the controller was in automatic mode for 90.6 +/- 9.6% of the time. The performance during automatic control showed the mean arterial pressure to be within 10 mmHg of the setpoint for 71.4 +/- 15.5% of the time. The average absolute distance to the setpoint was 8.1 +/- 7.2 mmHg. CONCLUSIONS: The overall performance of the controller was noted as very satisfactory by the anesthesiologists.

Evaluation of automatically learned intelligent alarm systems.

In this contribution it is investigated whether a combination of mathematical simulation and inductive machine learning can replace the usual knowledge elicitation techniques. To test this a domain was selected for which knowledge based systems had a high performance: intelligent alarm systems. A mathematical model of a breathing circuit and ventilated patient was implemented in PSpice. Airway pressure, gas flows and CO2 concentration were simulated with this model, during normal functioning of the breathing circuit and during several mishaps, for a wide range of simulated patients. With an inductive machine learning program, classification trees were created from the simulated patient data. The classification trees described each breathing circuit mishap in terms of changes in signal feature values with respect to the normal situation and were implemented as alarm system knowledge bases. The alarm systems were tested with data measured at 17 mechanically ventilated animals. During ventilation of the animals several mishaps were introduced. For each animal, 93-100% of all mishaps could be detected correctly by the alarm systems. The false alarm rate ranged on average from one false alarm per h to one false alarm every 2.5 h. It was concluded that the suggested approach to knowledge elicitation was successful.

Temporal logics and real time expert systems.

This paper introduces temporal logics. Due to the eternal compromise between expressive adequacy and reasoning efficiency that must decided upon in any application, full (first order logic or modal logic based) temporal logics are frequently not suitable. This is especially true in real time expert systems, where a fixed (and usually small) response time must be guaranteed. One such expert system, Fagan's VM, is reviewed, and a delineation is given of how to formally describe and reason with time in medical protocols. It is shown that Petri net theory is a useful tool to check the correctness of formalised protocols.

Correction for respiration artifact in pulmonary blood pressure signals of ventilated patients.

OBJECTIVE: To develop an algorithm that corrects pulmonary artery pressure signals of ventilated patients for the respiration artifact. The algorithm should test the validity of the pulmonary pressure signal and differentiate between the cyclic respiration artifact and true measurement artifacts. METHODS: The shape of each pulmonary pressure beat is described by eight characteristic features, including mean pressure value and the systolic and diastolic timing and pressure values. The features are corrected for the respiration artifact by fitting them in a least-squares sense on the first and second harmonics of the ventilator frequency. The corrected features are used by a signal validation algorithm, which adds a validity flag to each pressure beat. The validation algorithm rejects pressure beats with sudden changes in their shape but adapts itself when the changes persist. RESULTS: The performance of the correction and validation technique was evaluated using pulmonary artery pressure signals of 30 patients who were scheduled for open heart surgery. The algorithm correctly recognized as invalid data those pressure signals disturbed by coagulation, surgical manipulations, or flushes of the pressure line. The algorithm marked on average 77 +/- 11% of the pulmonary pressure beats as valid. CONCLUSIONS: The validation algorithm marked sufficient pressure beats as valid to update a trend display every 5 sec. The correction algorithm enabled the validation algorithm to differentiate between true measurement artifacts and the respiration artifact.

Building intelligent alarm systems by combining mathematical models and inductive machine learning techniques Part 2--sensitivity analysis.

In an earlier study an approach was described to generate intelligent alarm systems for monitoring ventilation of patients via mathematical simulation and machine learning. However, ventilator settings were not varied. In this study we investigated whether an alarm system could be created with which a satisfactory classification performance could be obtained under a wide variety of ventilator settings, by varying inspiratory to expiratory time (I:E) ratio, tidal volume and respiratory rate. In a first experiment three patient data sets were modeled, each with a different I:E ratio. A part of each data set was used to construct an alarm system for each I:E ratio. The remaining part was used to test the performance of the alarm systems. The three training sets were also combined to construct one alarm system, which was tested with the three test sets. Finally, all alarm systems were tested with data generated by a patient simulator. Similar experiments were performed for the tidal volume and the respiratory rate. It was concluded that an optimally functioning alarm system should contain a library of rule sets, one for each set of ventilator settings. A second best alternative is to take all possible settings into consideration when constructing the training set. Classification performance of the trees that were trained with multiple ventilator settings ranged from 98 to 100% for all test sets. When tested with the independent patient simulator data the classification performance of these trees ranged from 80 to 100%.

OpenLabs services for ordering laboratory investigations.

An automated system is described that screens requests for laboratory investigations from GPs and delivers feedback with respect to the adequacy of the requests. The system has to replace and extend the current system in which feedback is provided on a manual basis each half year, based on the tests requested during a period of one month, randomly selected from the previous half year. It has been reported elsewhere that the manual system reduced the number of tests requested considerably. The criteria used by the automated system and the manual system are based on guidelines and work agreements that GPs have agreed to follow when requesting investigations. It is concluded that the automated system is very user-friendly and that in the order of 4-17% of the requested tests could be identified as unnecessary, with a false negative rate in between 4% (for hyperthyroidism) and 23% (for hypothyroidism). The achieved reduction in the number of tests is in addition to the reduction obtained in the manual system.

Automated infusion of nitroglycerin to control arterial hypertension during cardiac surgery.

OBJECTIVE: To evaluate the feasibility of closed-loop blood pressure control during cardiac surgery. DESIGN: A closed-loop system regulated peroperative hypertension by controlling the infusion rate of the vasodilator nitroglycerin (NTG). The controller consisted of a regulator which was monitored by a supervisory computer program. Mean arterial pressure (MAP) was calculated every 5 s from measurements of the radial artery pressure signal. The regulator calculated an NTG infusion rate with each new MAP measurement. The supervisory computer program monitored the regulator's actions and adapted or overruled the regulator when required. SETTING: The cardiac surgery operating room. PATIENTS: 46 patients who were scheduled for cardiac surgery and who developed peroperative hypertension. INTERVENTIONS: Patients were scheduled for either bypass or valve replacement surgery. The closed-loop system was used to control hypertension before and after cardiopulmonary bypass. The use of the closed-loop system did not require deviation from the protocol normally used during cardiac surgery. All patients received standard continuous anaesthesia with opioids. MEASUREMENTS AND RESULTS: Initial automatic control was achieved in 9.4 (4.1 SD) min. The percentage of time that MAP remained in a range around the target MAP of +/- 10 and +/- 20 mmHg was 74 and 94%, respectively. The mean NTG infusion rate while MAP was within 5 mmHg of target MAP was 1.14 (0.84 SD) micrograms kg-1 min-1. Target MAP was set between 65 and 90 mmHg. There was a small group of patients (6 out of 46) who did not respond to NTG and required alternative drug therapy. CONCLUSIONS: The controller provided fast and stable control in all patients. The expert knowledge implemented through the supervisory computer program enabled the controller to respond adequately to the rapid changes in arterial pressures commonly associated with cardiac surgery. We conclude that closed-loop control of arterial pressure is feasible not only in the cardiac surgical care unit but also during cardiac surgery.

Building intelligent alarm systems by combining mathematical models and inductive machine learning techniques.

In this article a technique is described to develop knowledge-based alarm systems for ventilator therapy, using mathematical modeling and machine learning. With a mathematical model airway pressure, expiratory gas flow and CO2 concentration at the endotracheal tube are simulated for patients, undergoing volume-controlled ventilation with constant ventilator settings, during normal functioning of the breathing circuit and during breathing circuit mishaps (leaks and obstructions). Simulations were performed for 94 physiologically different 'patients', by varying airway resistance and lung/thorax compliance values in the model. Each simulated breath was described by a set of derived signal features and a label that constituted during which event (normal function or mishap) the breath was recorded. With an inductive machine learning algorithm rules, linking signal feature values to breathing circuit events, were created from data of 54 of the simulated patients. The resulting set of rules was able to classify 99% of events in the data of the remaining 40 patients correctly. Of signals, measured at a ventilated lung simulator, 100% of events were classified correctly.

Optimal surface electrode positioning for reliable train of four muscle relaxation monitoring.

In the clinic, a major problem in train of four (TOF) muscle relaxation monitoring is incorrect placement of stimulation and recording electrodes, frequently resulting in incorrect estimates of the patient's degree of relaxation or in abandonment of relaxation monitoring. The aim of this study was to arrive at recommendations that describe how to find optimal positions for the electrodes, where 'optimal' is taken in the sense that small deviations from these positions introduce no or only a small decline in the accuracy of the computed degree of muscle relaxation. This study, which employed the Relaxograph as the stimulation and measuring device, established that incorrect positioning is a real problem that frequently occurs; that the correctness of positioning is not guaranteed when the calibration of the Relaxograph succeeds; that the inadequacy of the electrode position is sometimes discovered for the first time when relaxation deepens; that positioning errors can be discovered by analysing the shape of the evoked compound action potential (ECAP), not only upon calibration but also when relaxation deepens; that a set of optimal electrode positions can be found; and that recommendations of how to find these optimal positions could help clinicians to place the electrodes in such a way, that reliable relaxation monitoring was possible in 100% of the investigated cases. In a first test in 30 adult patients, we surveyed how clinicians routinely positioned electrodes and found that in 14 of the 30 cases positioning was unsuccessful. In a second test in 10 patients, we tested a variety of electrode positions in order to discover 'optimal' stimulation, recording and ground electrode sites. In a third test in 10 patients, electrodes were positioned at these 'optimal' sites; stimulation and recording at these sites was successful in all 10 cases.