Machine Learning and Decision Support in Critical Care

Abstract

Clinical data management systems typically provide caregiver teams with useful information, derived from large, sometimes highly heterogeneous, data sources that are often changing dynamically. Over the last decade there has been a significant surge in interest in using these data sources, from simply re-using the standard clinical databases for event prediction or decision support, to including dynamic and patient-specific information into clinical monitoring and prediction problems. However, in most cases, commercial clinical databases have been designed to document clinical activity for reporting, liability and billing reasons, rather than for developing new algorithms. With increasing excitement surrounding "secondary use of medical records" and "Big Data" analytics, it is important to understand the limitations of current databases and what needs to change in order to enter an era of "precision medicine." This review article covers many of the issues involved in the collection and preprocessing of critical care data. The three challenges in critical care are considered: compartmentalization, corruption, and complexity. A range of applications addressing these issues are covered, including the modernization of static acuity scoring; on-line patient tracking; personalized prediction and risk assessment; artifact detection; state estimation; and incorporation of multimodal data sources such as genomic and free text data.

Keywords: Critical care; feature extraction; machine learning; signal processing.

Fig 1

Overview of the primary challenges in critical care. The three challenges that are presented to researchers in this field are discussed in turn: the compartmentalization of the data, which results in disparate datasets that are difficult to acquire and interrogate; the corruption of the data during collection, which necessitates non-trivial corrective work; and the complexity inherent in the systems monitored.

Fig 2

Example of a false alarm which incorrectly asserted the patient was in asystole. The signals shown are the photoplethysmogram (PPG, top in green), the electrocardiogram lead V (ECG, middle in blue), and the electrocardiogram lead II (ECG, bottom in red). The alarm likely triggered univariately on ECG lead V. At least two methods reviewed in this section could have prevented this false alarm: the use of signal quality on lead V or a multimodal data fusion approach which incorporated ECG lead II, the PPG, or both.

Fig 3

Example of low, sometimes zero respiratory rates. As a sustained breathing rate of zero for hours is incompatible with life, the data here may represent: i) undersampling of true respiratory distress with intermittent apnea, ii) erroneous data corresponding to sensor fault, or iii) manually entered data intended to represent poor physiologic state.

Fig 4

Example of a Gaussian Process (GP) regression inferring the value of missing data on an unevenly sampled time series of hematocrit values. The raw values are plotted as red circles against the mean of the GP (solid green line) and the 95% confidence intervals (dashed green lines).

Fig 5

Supervised learning in dynamic Bayesian networks. Graphical model representation of the switching vector autoregressive (switching VAR) is depicted in panel (a). Panels (b) shows the unrolled representation (with respect to time and inference steps) of the two models, with an added logistic regression layer (elliptic nodes) which utilize the marginals over the discrete latent variables as features for time-series classification [an example of inferred marginals is shown at the bottom of the panel (b)]. These unrolled structures, which resemble recurrent neural networks, allow for efficient supervised learning and inference via error backpropagation.