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

Enhanced notification of critical ventilator events

Mechanical ventilators are designed to generate alarms when patients become disconnected or experience other critical ventilator events. However, these alarms can blend in with other accustomed sounds of the intensive care unit. Ventilator alarms that go unnoticed for extended periods of time often result in permanent patient harm or death. We developed a system to monitor critical ventilator events through our existing hospital network. Whenever an event is identified, the new system takes control of every computer in the patient's intensive care unit and generates an enhanced audio and visual alert indicating that there is a critical ventilator event and identifies the room number. Once the alert is acknowledged or the event is corrected, all the computers are restored back to the pre-alert status and/or application. This paper describes the development and implementation of this system and reports the initial results, user acceptance, and the increase in valuable information and patient safety.

Do we need a systematic activation of alarm soundings for blood pressure monitoring for the safety of ICU patients?

PURPOSE

Evaluation of a protocol of activation-deactivation of alarm soundings on noninvasive mean arterial pressure (MAP) monitoring.

MATERIALS AND METHODS

103 patients were analyzed. Alarm soundings on noninvasive MAP were either activated (group A) or inactivated (group NA) according to a protocol based on patient's hemodynamic status.

RESULTS

In the first part of the study (11 periods of 24 hours), we informed the nursing staff before data collection. In the second part of the study (11 periods of 24 hours), we did not inform the nursing staff before data collection. Results of the two parts were assembled and the final analysis consisted of 3,304 hours of monitoring with 12,839 noninvasive MAP measurements, 6,354 in group A and 6,485 in group NA. In group A, 75 of 944 (8%) of abnormal measurements were detected or managed with a delay > 15 minutes and among them 16 required a therapeutic adjustment. In group NA, 140 of 265 (53%) of abnormal measurements were detected or managed with a delay > 15 minutes and among them, only one required a therapeutic adjustment. The number of alarm soundings avoided in group NA was 286.

CONCLUSION

In this study, a systematic activation of alarm soundings for non-invasive blood pressure monitoring seemed unnecessary for the safety of ICU patients.

Analysis of respiratory pressure-volume curves in intensive care medicine using inductive machine learning

We present a case study of machine learning and data mining in intensive care medicine. In the study, we compared different methods of measuring pressure-volume curves in artificially ventilated patients suffering from the adult respiratory distress syndrome (ARDS). Our aim was to show that inductive machine learning can be used to gain insights into differences and similarities among these methods. We defined two tasks: the first one was to recognize the measurement method producing a given pressure-volume curve. This was defined as the task of classifying pressure-volume curves (the classes being the measurement methods). The second was to model the curves themselves, that is, to predict the volume given the pressure, the measurement method and the patient data. Clearly, this can be defined as a regression task. For these two tasks, we applied C5.0 and CUBIST, two inductive machine learning tools, respectively. Apart from medical findings regarding the characteristics of the measurement methods, we found some evidence showing the value of an abstract representation for classifying curves: normalization and high-level descriptors from curve fitting played a crucial role in obtaining reasonably accurate models. Another useful feature of algorithms for inductive machine learning is the possibility of incorporating background knowledge. In our study, the incorporation of patient data helped to improve regression results dramatically, which might open the door for the individual respiratory treatment of patients in the future.

[Selective activation of blood pressure monitoring alarms: effect of noise pollution in the intensive care unit]

OBJECTIVE

To evaluate a selective activation of sounding alarms on non-invasive blood pressure (BP) monitoring according to the patient haemodynamic status.

STUDY DESIGN

Prospective study.

METHODS

Activation of alarms on BP was regulated with a protocol. Sounding alarms were either inactivated when patient's haemodynamic status was stable (group 1), or activated when it was unstable (group 2). The frequency of BP measurement was one every 15 min. For all mean BP value recorded, the following criteria were analyzed: 1) normality of the value compared to ranges 65-115 mmHg in group 1 or compared to alarm thresholds in group 2; 2) consequences on the care and therapeutic; 3) delay when an abnormal value was detected and managed after more than 15 min.

RESULTS

1,674 hours of monitoring from 42 patients, allowed the analysis of 6,695 measurements of mean BP, 3,092 in group 1 and 3,603 in group 2. In group 1, 2,822 measurements were considered as normal and 3,094 measures in group 2. Eight measurements had consequences on therapeutic in group 1, with only one with delay in care giving. 287 measurements had consequences on therapeutic in group 2, 8 with delay in care giving. Six per cent of abnormal measurements in group 2 were managed with delay. This protocol reduced by 52% the production of sounding alarms on BP, without noxious effects for the patients.

CONCLUSION

Selective activation of sounding alarms on BP, according to the patient haemodynamic status, reduced noise pollution and could be one solution to improve monitoring efficiency in intensive care unit.

[Clinical evaluation of alarm efficieny in intensive care]

OBJECTIVE

To evaluate the efficiency of haemodynamic and respiratory monitoring system by a clinical analysis of the alarms.

STUDY DESIGN

Observational prospective study.

PATIENTS

25 patients who presented acute respiratory distress syndrome and who were monitored with haemodynamic and respiratory monitoring.

METHODS

Each minute, a bedside clinical observer analysed alarms from the monitoring according to detection or absence to clinical events. Four situations were defined to statistical descriptive analysis: a) false positive (FP); b) true positive (TP); c) false negative (FN); and d) true negative (TN). True positive alarm which induced consequences on patients care were also analysed.

RESULTS

15,013 minutes allowed the recordings of 3,665 alarms, 44% from arterial pressure, 17% from SpO2 and 12% from airways maximal pressure. 46% were false positive alarms inducing a noisy pollution. The positive predictive value PPV = TP/(TP + FP) of these alarms were respectively 51% for arterial pressure, 18% for SpO2 and 100% for Paw. Only 5% of true positive alarms induced consequences on patients care.

CONCLUSION

This protocol allowed the evaluation of monitoring efficiency. This kind of evaluation may help to improve monitoring capacity with reducing noisy pollution from false positive alarms.

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.

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%.