Fabregat A, Magret M, Ferré JA, Vernet A, Guasch N, Rodríguez A, Gómez J, Bodí M
A Machine Learning decision-making tool for extubation in Intensive Care Unit patients
Comput Methods Programs Biomed 2021;200:105869
Invasive Mechanical Ventilation (IMV) is central to treating patients who are unable
to maintain adequate pulmonary ventilation and oxygenation to allow patients to recover.
Although IMV can be a life-saving procedure, it also bears significant risks such
as ventilator-induced lung injuries or infections as well as long-term problems after
recovery. One of the critical decisions regarding IMV is weaning. This includes, amongst
other steps, the removal of the endotracheal tube. Patients that need to be reintubated
bear several risks and problems associated, including increased mortality (25%-50%).
The goal of the current work was to create a machine learning (ML) model that can
increase the successful extubation rate in adult Intensive Care Unit (ICU) patients.
The model is based on data routinely collected from patients’ health record data.
Patients included were admitted to an ICU in a Spanish hospital between 2015 and 2019,
and received at least 12 consecutive hours of IMV. Variables used for prediction were
categorized in four types: T1: time series data (averaged over 20 minutes) (e.g.,
heart rate); T2: derived variables from T1 (e.g., respiratory rate); T3: discrete
event information, (e.g., Glasgow coma scale (GCS)); T4: demographics and admission
information (e.g., gender). In total, 20 predictors were used. The resulting dataset
had a strong imbalance with regard to successful extubations (1,108 versus 100). Therefore,
randomly selected data points of the most frequent class (successful extubation) were
removed for the training data set and/or a weight was assigned to data points. Seven-fold
cross-validation was determined appropriate. Extubation/Reintubation was basically
identified by finding gaps on the IMV monitor signal larger than 48 hours. As several
possible errors have an influence on this gap, a comparison with medical records was
necessary to correct the numbers (final dataset: 647 successful and 50 failed). Three
different ML classifiers were compared: support vector machine (SVM) with radial basis,
gradient boosting machine (GBM) with Bernoulli loss, and Linear Discriminant Analysis
(LDA). Mean Accuracy and AUROC were used to determine performance. The following scores
were achieved: SVM 94.6% and 98.3%; GBM 87% and 96%; LDA 72% and 79%. The results
suggest that the top five predictors in descending order of importance are time, GCS,
body mass index, respiratory rate-oxygenation index, and plateau pressure. On the
other hand, the least relevant predictors in descending order of importance are Spanish
Society of Intensive, Critical Medicine and Coronary Units classification code for
ICU admission reason, gender, total cumulative dose, total given dose, and ventilation
mode. The models should not be applied as a general-purpose predictor of success for
programmed extubations or as a monitoring alarm system but as a support tool to validate
the medical staff's decision. With the predictive accuracy achieved, the rate of failed
extubation (currently 9%) could be reduced to a theoretical 1%. The results suggest
that ML tools are especially well suited to support the decision-making protocol based
on spontaneous breathing trials to decide about extubation.
Kempa-Liehr AW, Lin CYC, Britten R, Armstrong D, Wallace J, Mordaunt D, O'Sullivan
M
Healthcare pathway discovery and probabilistic machine learning
Int J Med Inform 2020;137:104087
The success of electronic health records has also driven several other research areas
such as knowledge management in healthcare, which basically involves four steps: (1)
data access; (2) knowledge discovery; (3) knowledge translation and interpretation,
as well as (4) knowledge description, integration and sharing. An important role hereby
is played by healthcare pathways that incorporate the operational knowledge of a healthcare
organization by defining the execution sequence of clinical activities as patients
move through a treatment process. In many cases, these pathways result from clinician-led
practice rather than explicit design, which leads to several problems (e.g., lack
of update). The study aims to combine healthcare pathway discovery with predictive
models of individualized recovery times after appendicectomy. Particular emphasis
is set on easy to interpret models for clinicians. The predictive model takes the
stochastic volatility of pathway performance indicators into account and can replicate
the dominant mode as well as the fat tail of the empirical recovery time distribution.
To mine the pathways, the ProM software was used. First, healthcare pathway variations
were discovered and then reduced (clustering, merging consecutive activities, condense
repetitive patterns) to meaningful models. In a second step, conformance of these
models with actual patient traces is evaluated, including new findings into the model
leads to an iterative approach between pathway discovery and conformance analysis.
The third step involves data enrichment, which comprises two stages: healthcare pathway
performance evaluation and healthcare pathway performance analysis. The main objectives
of evaluating healthcare pathway performance are to understand the strengths and weaknesses
of the current pathway design. Analyzing the performance of healthcare pathways with
respect to pathway variants and other possible influencing factors like demographics
or patient-specific pathway characteristics (e.g., surgery duration) is the final
step of the proposed process mining pipeline. For the appendicitis model, 13 pathway
variants were discovered, whereas the top four variants accounted for approximately
88% of the patient traces. In a next step, it was analyzed if the variants are relevant
features or covariates for explaining the stochastic volatility of postoperative length
of stay. To build two probabilistic machine learning models, 415 individual patient
traces were used. The two models showed promising results to explain the length of
stay. Summarizing, the proposed process mining pipeline successfully constructed concise
pathway models for the appendicitis case study and, therefore, supported the generation
of probabilistic machine learning models.
Li Y, Nair P, Lu XH, Wen Z, Wang Y, Dehaghi AAK, Miao Y, Liu W, Ordog T, Biernacka
JM, Ryu E, Olson JE, Frye MA, Liu A, Guo L, Marelli A, Ahuja Y, Davila-Velderrain
J, Kellis M
Inferring multimodal latent topics from electronic health records
Nat Commun 2020;11(1):2536
Electronic health records (EHRs) are heterogeneous collections of patient health information
that would support multiple uses such as risk prediction, clinical recommendations,
or individual therapeutic concepts. However, raw data in EHRs is in many cases not
directly processable, especially when building formal models. Different challenges
such as non-standardized clinical notes, heterogeneous data types, missing standardization,
or diagnosis-driven lab tests pose challenges. Appropriate and effective computational
methods have the potential to overcome those challenges and provide access to an encyclopedia
of diseases, disorders, injuries, and other related health conditions, uncovering
a modular phenotypic network. The paper introduces MixEHR to: (1) distill meaningful
disease topics from otherwise highly sparse, biased, and heterogeneous EHR data; and
(2) provide clinical recommendations by predicting undiagnosed patient phenotypes
based on their disease mixture membership. MixEHR builds on collaborative filtering
and latent topic modeling and can model various EHR categories with separate discrete
distributions. A variational inference algorithm that scales to large-scale EHR data
was created. The model was applied to three EHR datasets: (1) Medical Information
Mart for Intensive Care (MIMIC)-III (50,000 intensive care unit admissions); (2) Mayo
Clinic EHR dataset containing 187 patients, including with 93 bipolar disorders and
94 controls; (3) The Régie de lʼassurance maladie du Québec Congenital Heart Disease
Dataset (Quebec CHD Database; more than 80,000 patients with congenital heart disease).
The authors followed a probabilistic joint matrix factorization approach. The high
dimensional and heterogeneous clinical record was projected onto a low dimension probabilistic
meta-phenotype signature, reflecting the patient's mixed memberships across diverse
latent disease topics. Factorization is carried out at two levels. At the lower level,
data-type-specific topic models, learning a set of basis matrices for each data type,
were applied. A common loading matrix that connects the multiple data types for each
patient was used at the higher level. The approach was used, among others, to define
a disease comorbidity network, create patient risk prioritization, EHR code predictions,
or mortality predictions from the given datasets. Overall, the MixEHR approach's accuracy
scores top compared to other existing approaches. MixEHR can infer expected phenotypes
of a patient conditioned only on a subset of clinical variables that are perhaps easier
and cheaper to measure. Currently, data are a set of two-dimensional matrices of patients
by measurements in the model. To model higher dimensional objects such as patient
by lab test by diagnoses, MixEHR could be extended to a probabilistic tensor-decomposition
framework.
Weemaes M, Martens S, Cuypers L, van Elslande J, Hoet K, Welkenhuysen J, Goossens
R, Wouters S, Houben E, Jeuris E, Jeuris K, Laenen L, Bruyninckx K, Beuselinck K,
André E, Depypere M, Desmet S, Lagrou K, Van Ranst M, Verdonck AKLC, Goveia J
Laboratory information system requirements to manage the COVID-19 pandemic: A report
from the Belgian national reference testing center
J Am Med Inform Assoc 2020;27(8):1293–9
The paper describes the challenges faced by the Belgian National Reference Center
for COVID-19 testing at the University Hospitals Leuven, when demand passed allocated
surge capacity during the initial phases of the COVID-19 pandemic. This includes the
design, implementation and requirements of laboratory information system (LIS) functionality
related to managing increased test demand during the COVID-19 crisis. In particular,
all phases in laboratory testing were streamlined: the pre-laboratory phase (test
ordering, sample packaging, and shipping); the pre-analytical phase (sample registration,
tracking, and test prioritization); and the post-analytical phase (automated reporting
and facilitating data-driven policy-making). Apart from COVID-19 testing, the laboratory
concerned performs more than 12,000,000 lab tests a year. The LIS is in-house developed
and maintained by a dedicated team. The system includes a computerized physician order
entry (CPOE) module for in-house test ordering, which is fully integrated into the
electronic health record (EHR). All external orders were initially paper-based and
required that request forms accompany the sample. In the course of the analysis, 17
major challenges were identified in the different phases of the testing process. Selected
solutions included: a COVID-19 specific CPOE module was linked to both the LIS and
EHR, allowing to automatically retrieve demographic information, which dramatically
improved metadata completeness; a “COVID-19 status” button on the main page of the
EHR of each patient was displayed, showing in real-time the results of SARS-CoV-2
laboratory testing; a database with contact details and preferred reporting methods
(e.g., fax, email, electronic mailbox system) of every laboratory in Belgium was compiled,
to enable automated test reporting (resulted in more than 98% automated reporting).
To successfully implement such changes in a short time, several prerequisites apply.
The authors, therefore, recommend that crisis management teams not only consist of
staff focused on increasing analytical capacity but also information technology-staff
and to apply change management frameworks. To summarize, the most effective solutions
reported were to streamline sample ordering through a CPOE system and reporting by
developing a database with contact details of all laboratories in Belgium. In addition,
the implementation of R/Shiny-based statistical tools facilitated epidemiological
reporting and enabled explorative data mining.
