The deterioration of organ function in ICU patients requires swift response to prevent further damage to vital systems. Focusing on the circulatory system, we build a model to predict if a patient's state will deteriorate in the near future. We identify circulatory system dysfunction using the combination of excess lactic acid in the blood and low mean arterial blood pressure or the presence of vasoactive drugs. Using an observational cohort of 45,000 patients from a Swiss ICU, we extract and process patient time series and identify periods of circulatory system dysfunction to develop an early warning system. We train a gradient boosting model to perform binary classi cation every ve minutes on whether the patient will deteriorate during an increasingly large window into the future, up to the duration of a shift (8 hours). The model achieves an AUROC between 0.952 and 0.919 across the prediction windows, and an AUPRC between 0.223 and 0.384 for events with positive prevalence between 0.014 and 0.042. We also show preliminary results from a recurrent neural network. These results show that contemporary machine learning approaches combined with careful preprocessing of raw data collected during routine care yield clinically useful predictions in near real time.
Despite the high level of monitoring in the ICU, it is often infeasible for doctors
to continually monitor the state of all patients. Unanticipated deteriorations can
be life-threatening and require swift response. Identifying imminent or likely
deterioration in a timely fashion is therefore an important question[
In this work, we describe a data-driven predictive model for circulatory system failure. We integrate continuous measurements from hundreds of physiological variables and treatment parameters, drawn from a dataset of 44,655 patients over 8 years comprising 553.18 years of patient data. Our system identi es patterns indicative of pending haemodynamic instability, using many more variables than a typical ICU physician could assess. Currently relying on an observational dataset for internal validation, once nalised, this system will be deployed in a Swiss ICU to clinically validate its use as a real-time monitoring system. 1.1
Related Work
Risk strati cation on the basis of physiological parameters is a common practice
in the ICU, and scores such as SOFA[
Preparing the data for use in a machine learning system was a critical component of this work. Raw data was exported from the patient database management system deployed at Bern University Hospital and then processed in several steps. Routinely-collected data of this kind features many challenges for computational analysis. Errors in data labelling (for example venous versus arterial blood gases), missing and implausible values, ambiguous or contradictory records, as well as artefacts introduced by routine care (for example blood pressure spikes due to arterial line ushing) necessitate care during data processing. In this work we attempted to remove the suspected erroneous data using variable-speci c processing, resolving or deleting ambiguous records, and removing values based on plausible physiological ranges.
To deal with missing data, we compute the median sampling interval for each variable from training data, and forward- ll up to this point. After this, we decay to a rolling local median from the recent past (calculated over a similar interval). This re ects the belief that frequently-measured variables vary rapidly and should not be forward- lled for long, while decaying to the recent median value implies that in the absence of data, we assume the patient has returned to `baseline' (for them), where this can vary throughout their stay.
Medications were converted from doses to ow rates, treating `instantaneous' drugs (such as tables) as ows over an e ective active period, which we de ned for each drug. In the database system used by the ICU at Bern University Hospital, drugs often received multiple unique identi ers for di erent dosage options, corresponding to di erent variables in the database. To address this and other redundancies in the data (for example, three ways of measuring temperature) we performed a manual dimensionality reduction step, merging variables that we identi ed to be su ciently similar. In doing so, we reduced the total number of variables from 728 to 209. This means the model is applicable in any system measuring these variables, and is not speci c to the ICU in Bern.
We use pandas[
We de ne deterioration as the appearance of a `worse' state during a window
up to t hours in the future. A patient can be in one of four states, where
0 is the best (stable), and states 1-3 describe increasing levels of circulatory
system dysfunction. This dysfunction is identi ed through impaired circulatory
function and elevated lactate values ( 2 mmol/L). Impaired circulatory function
requires either low ( 65 mmHg) mean arterial pressure (MAP) or the presence
of vasoactive drugs. To minimize spurious calls[
The three dysfunctional levels are de ned by the type and intensity of vasoactive drugs: level drugs requirement 1 Any dose of dobutamine, milrinone, levosimendan, or theophylline 2 < 0:1 g/kg/minute of norepinephrine or epinephrine 3 0:1 g/kg/minute of norepinephrine or epinephrine, or any dose of vasopressin The task is then binary classi cation on whether a patient in state s at time t will be in a state s + s ( s > 0) during a window starting ve minutes from t and ending t + t hours later. We consider t in increments of one hour up to eight hours, the duration of a shift. For t = 8, this means the model can ag patients who may need additional attention during the next shift, but who may not be imminently critical.
As a model, we use an ensemble approach of boosted decision trees in the
LightGBM library[
In our experimental setup, we use the most recent six months of data to construct the test set, re ecting that such models are necessarily trained on retrospective data, and will be applied on new patients with a slightly di erent data distribution. We report AUROC as well as area under the precision-recall curve as deteriorations are relatively rare (1.4% prevalence during a one-hour window). The results for varying t are shown in Figure 3. We see that AUROC remains high as t increases, indicating high accuracy for predictions over the next shift. AUPRC is more challenging for this model and task, although performance is well above baseline (dotted line) for all t (between 9.14x and 15.93x). Given the preliminary nature of the LSTM results and the limited input data it receives (at most four hours), its performance is promising.
1.0 0.9 C0.8 O R U0.7 A 0.6 We have shown how careful handling of a large retrospective cohort of ICU patients results in a predictive model of circulatory organ failure with high (AUROC > 0.9) performance on time-horizons up to 8 hours in the future. We are currently developing models based on recurrent neural networks to better exploit the temporal nature of this data, and studying the behaviour of our trained classi er to identify ways to enhance positive predictive value. One direction is to provide the recurrent neural networks with longer history of data, such as the entire patient stays, so that the models can use as much information as available to make more accurate predictions. Once satisfactory in silico, this model will be deployed in the ICU for external validation. This work demonstrates the potential for large-scale multivariate modelling to identify patterns in physiological signals, enabling early warning of circulatory system deterioration.