Early Prediction of Cardiogenic Shock Using Machine Learning

Cardiogenic shock (CS) is a severe condition with in-hospital mortality of up to 50%. Patients who develop CS may have previous cardiac history, but that may not always be the case, adding to the challenges in optimally identifying and managing these patients. Patients may present to a medical facility with CS or develop CS while in the emergency department (ED), in a general inpatient ward (WARD) or in the critical care unit (CC). While different clinical pathways for management exist once CS is recognized, there are challenges in identifying the patients in a timely manner, in all settings, in a timeframe that will allow proper management. We therefore developed and evaluated retrospectively a machine learning model based on the XGBoost (XGB) algorithm which runs automatically on patient data from the electronic health record (EHR). The algorithm was trained on 8 years of de-identified data (from 2010 to 2017) collected from a large regional healthcare system. The input variables include demographics, vital signs, laboratory values, some orders, and specific pre-existing diagnoses. The model was designed to make predictions 2 h prior to the need of first CS intervention (inotrope, vasopressor, or mechanical circulatory support). The algorithm achieves an overall area under curve (AUC) of 0.87 (0.81 in CC, 0.84 in ED, 0.97 in WARD), which is considered useful for clinical use. The algorithm can be refined based on specific elements defining patient subpopulations, for example presence of acute myocardial infarction (AMI) or congestive heart failure (CHF), further increasing its precision when a patient has these conditions. The top-contributing risk factors learned by the model are consistent with existing clinical findings. Our conclusion is that a useful machine learning model can be used to predict the development of CS. This manuscript describes the main steps of the development process and our results.

Early prediction of hemodynamic interventions in the intensive care unit using machine learning

Background

Timely recognition of hemodynamic instability in critically ill patients enables increased vigilance and early treatment opportunities. We develop the Hemodynamic Stability Index (HSI), which highlights situational awareness of possible hemodynamic instability occurring at the bedside and to prompt assessment for potential hemodynamic interventions.

Methods

We used an ensemble of decision trees to obtain a real-time risk score that predicts the initiation of hemodynamic interventions an hour into the future. We developed the model using the eICU Research Institute (eRI) database, based on adult ICU admissions from 2012 to 2016. A total of 208,375 ICU stays met the inclusion criteria, with 32,896 patients (prevalence = 18%) experiencing at least one instability event where they received one of the interventions during their stay. Predictors included vital signs, laboratory measurements, and ventilation settings.

Results

HSI showed significantly better performance compared to single parameters like systolic blood pressure and shock index (heart rate/systolic blood pressure) and showed good generalization across patient subgroups. HSI AUC was 0.82 and predicted 52% of all hemodynamic interventions with a lead time of 1-h with a specificity of 92%. In addition to predicting future hemodynamic interventions, our model provides confidence intervals and a ranked list of clinical features that contribute to each prediction. Importantly, HSI can use a sparse set of physiologic variables and abstains from making a prediction when the confidence is below an acceptable threshold.

Conclusions

The HSI algorithm provides a single score that summarizes hemodynamic status in real time using multiple physiologic parameters in patient monitors and electronic medical records (EMR). Importantly, HSI is designed for real-world deployment, demonstrating generalizability, strong performance under different data availability conditions, and providing model explanation in the form of feature importance and prediction confidence.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13054-021-03808-x.

Inflation instability in the lung: an analytical model of a thick-walled alveolus with wavy fibres under large deformations

Inflation of hollow elastic structures can become unstable and exhibit a runaway phenomenon if the tension in their walls does not rise rapidly enough with increasing volume. Biological systems avoid such inflation instability for reasons that remain poorly understood. This is best exemplified by the lung, which inflates over its functional volume range without instability. The goal of this study was to determine how the constituents of lung parenchyma determine tissue stresses that protect alveoli from instability-related overdistension during inflation. We present an analytical model of a thick-walled alveolus composed of wavy elastic fibres, and investigate its pressure-volume behaviour under large deformations. Using second-harmonic generation imaging, we found that collagen waviness follows a beta distribution. Using this distribution to fit human pressure-volume curves, we estimated collagen and elastin effective stiffnesses to be 1247 kPa and 18.3 kPa, respectively. Furthermore, we demonstrate that linearly elastic but wavy collagen fibres are sufficient to achieve inflation stability within the physiological pressure range if the alveolar thickness-to-radius ratio is greater than 0.05. Our model thus identifies the constraints on alveolar geometry and collagen waviness required for inflation stability and provides a multiscale link between alveolar pressure and stresses on fibres in healthy and diseased lungs.

A rapidly deployable individualized system for augmenting ventilator capacity

Strategies to split ventilators to support multiple patients requiring ventilatory support have been proposed and used in emergency cases in which shortages of ventilators cannot otherwise be remedied by production or procurement strategies. However, the current approaches to ventilator sharing lack the ability to individualize ventilation to each patient, measure pulmonary mechanics, and accommodate rebalancing of the airflow when one patient improves or deteriorates, posing safety concerns to patients. Potential cross-contamination, lack of alarms, insufficient monitoring, and inability to adapt to sudden changes in patient status have prevented widespread acceptance of ventilator sharing. We have developed an individualized system for augmenting ventilator efficacy (iSAVE) as a rapidly deployable platform that uses a single ventilator to simultaneously and more safely support two subjects. The iSAVE enables subject-specific volume and pressure control and the rebalancing of ventilation in response to improvement or deterioration in an individual’s respiratory status. The iSAVE incorporates mechanisms to measure pulmonary mechanics, mitigate cross-contamination and backflow, and accommodate sudden flow changes due to subject interdependencies within the respiratory circuit. We demonstrate these capacities through validation using closed- and open-circuit ventilators on linear test lungs. We show that the iSAVE can temporarily ventilate two pigs on one ventilator as efficaciously as each pig on its own ventilator. By leveraging off-the-shelf medical components, the iSAVE could rapidly expand the ventilation capacity of healthcare facilities during emergency situations such as pandemics.

Cardiovascular system identification: Simulation study using arterial and central venous pressures

The paper presents a study of the identifiability of a lumped model of the cardiovascular system. The significance of this work from the existing literature is in the potential advantage of using both arterial and central venous (CVP) pressures, two signals that are frequently monitored in the critical care unit. The analysis is done on the system's state-space representation via control theory and system identification techniques. Non-parametric state-space identification is preferred over other identification techniques as it optimally assesses the order of a model, which best describes the input-output data, without any prior knowledge about the system. In particular, a recent system identification algorithm, namely Observer Kalman Filter Identification with Deterministic Projection, is used to identify a simplified version of an existing cardiopulmonary model. The outcome of the study highlights the following two facts. In the deterministic (noiseless) case, the theoretical indicators report that the model is fully identifiable, whereas the stochastic case reveals the difficulty in determining the complete system's dynamics. This suggests that even with the use of CVP as an additional pressure signal, the identification of a more detailed (high order) model of the circulatory system remains a challenging task.

Two-parameter leak estimation in non-invasive ventilation

In this paper we present a method for the estimation of leaks in non-invasive ventilation. Accurate estimation of leaks is a key component of a ventilator, since it determines the ventilator performance in terms of patient-ventilator synchrony and air volume delivery. In particular, in non-invasive ventilation, the patient flow is significantly different from the flow measured at the ventilator outlet. This is mostly due to the vent orifice along the tube that is used for exhalation, but also to the non-intentional leaks that occur elsewhere in the circuit (e.g., at the mask). Such leaks are traditionally quantified via a model with two parameters, but only one of them is continually updated - the other is fixed. The new algorithm allows for breath-by-breath update of both parameters. This was made possible by leveraging a model describing the patient respiratory mechanics.

[Vascular anomalies in the neonatal period]

OBJECTIVE

Vascular anomalies in the neonatal period are a diagnostic challenge for the lack of evident signs, symptoms and follow-up, and the convenience of restricting aggressive diagnostic tests. The aim of this work is to review the characteristics of neonatal cases presented to our Vascular Anomalies Unit in the last 5 years.

MATERIALS AND METHODS

All cases of suspected vascular anomaly presented to our unit before 1 month of age between 2010 and 2015 were reviewed, diagnostic tests and treatments carried out with chronology were analyzed. Presumptive diagnosis and final diagnosis (when available) were compared.

RESULTS

Fifteen vascular tumors were found, 2 with visceral involvement: 6 infantile hemangiomas (IH), 3 NICH, 4 RICH, 1 tufted hemangioma, 1 unspecified liver vascular tumor, 3 venous malformations (2 equivocal MRI and a hyperkeratotic venous malformation), 4 lymphatic malformations, 3 of them macrocystic, and 2 vascular lesions that were diagnosed of fibrosarcoma and sclerema neonatorum and they were not vascular anomalies. Only 3 patients with macrocystic lymphatic malformations had prenatal diagnosis.

CONCLUSION

Accurate diagnosis of vascular anomalies during the first month of life is difficult, even with MRI. Only in a few cases early treatment is needed, so it is worth taking time to follow-up. Different types of treatment (observation, propranolol, biopsy, laser, embolization, and resection) will depend on the condition to be treated. A continuous observation can avoid unnecessary procedures and risks.

End-inspiratory occlusion in the presence of intrinsic PEEP

Typical modern mechanical ventilators offer the clinician the possibility of automatically performing end-inspiratory occlusion maneuvers. The static conditions induced by such maneuvers are indeed favorable to estimating patient's respiratory mechanics in terms of total resistance (R) and elastance (E). These are parameters of wide clinical interest. However, in the presence of intrinsic PEEP, the standard formula used to compute E via the occlusion maneuver is known to be inaccurate. In this paper we propose an alternative method for the estimation of E via the occlusion maneuver that eliminates the bias by leveraging concepts derived from physiological modeling of the respiratory system dynamics. The proposed method is also capable of accounting for respiratory efforts triggering the breath, and hence can be applied in both passive and spontaneously breathing conditions.

Noninvasive estimation of alveolar pressure

This paper presents an algorithm for noninvasive estimation of alveolar pressure in mechanically ventilated patients who are spontaneously breathing. Continual monitoring of alveolar pressure is desirable to prevent ventilator-induced lung injury and to assess the intrinsic positive end-expiratory pressure (PEEPi), which is a parameter of clinical relevance in respiratory care and difficult to measure noninvasively. The algorithm is based on a physiological model of the respiratory system and, as such, it also provides insight into the respiratory mechanics of the patient under mechanical ventilation. In particular, the algorithm allows one to correctly estimate other clinical parameters of interest such as the patient's respiratory resistance and elastance, even in the presence of PEEPi.

Constrained optimization for noninvasive estimation of work of breathing

This paper presents a technique for noninvasive estimation of respiratory muscle effort (also known as work of breathing, WOB) in mechanically ventilated patients. Continual and real-time assessment of the patient WOB is desirable, as it helps the clinician make decisions about increasing or decreasing mechanical respiratory support. The technique presented is based on a physiological model of the respiratory system, from which a cost function is constructed as the sum of squared errors between model-based airway pressure predictions and actual measurements. Quadratic programming methods are used to minimize this cost function. An experimental example on animal data shows the effectiveness of the technique.

Allozyme, chloroplast DNA and RAPD markers for determining genetic relationships between Abies alba and the relic population of Abies nebrodensis

Allozyme, chloroplast (cpDNA) and random amplified polymorphic DNA (RAPD) markers have been used to estimate genetic and taxonomic relationships among different populations of Abies alba and the relic population of A. nebrodensis. Twelve isozyme gene loci, as well as restriction fragment length polymorphism (RFLP) at cpDNA spacer regions between t-RNA genes were analysed. Moreover, a set of 60 random sequence 10-mer primers were tested. Over all isozyme loci, evident differences in allele frequencies among A. nebrodensis and A. alba populations were found, particularly at 2 loci, phosphoglucose isomerase (Pgi-a) and shikimate dehydrogenase (Skd-a). More than 10% of the total genetic diversity was due to differences among populations. High values of genetic distances among populations were also found. Out of the 60 primers tested, 12 resulted in a polymorphic banding pattern both within and among populations. A total of 84 RAPD fragments were produced by the 12 selected primers. A phenogram of relationships among populations was constructed based on RAPD band sharing: the differentiation of the A. nebrodensis population was evident. The analysis of molecular variance (AMOVA) was used to apportion the variation among individuals within populations and among populations. There was considerable variation within each population: even so, genetic divergence was found among populations. This pattern of genetic variation was very different from that reported for inbred species. Identical cpDNA amplification and restriction patterns were observed among all the individuals sampled from the populations. Taken together, the results of allozyme and RAPDs show a clear differentiation among A. nebrodensis and A. alba populations and provide support for their classification into two different taxonomic groups.