8. Machine Learning & AI at the ICU
ML / ICU
Key factors:
⢠One of the most data dependent clinical environments is the critical care
department (CCD) in any of its forms: intensive care unit (ICU), pediatric
intensive care unit (PICU), neonatal intensive care unit (NICU) or surgical
inten sive care units (SICU), and this involves very practical implications for
MDSS at the point of care.
⢠Many of their patients are technologically dependent on the life-sustaining
devices that surround them. Beyond treatment, assessment of prognosis in
critical care and patient stratification combining different data sources is
extremely important in a patient-centric environment.
⢠The team supporting the patient ultimately must define what is required
and, in order to support clinical decision making, it is also necessary to
include other data from the electronic health record and monitoring
devices. These include fluid intake and patient output, demographic
information, laboratory blood draw assessments, medical images, and so on.
9. Machine Learning & AI at the ICU
ML / ICU
Key factors:
⢠Fresh approaches to data analysis tailored to the needs of the ICU
environments are required, and some of the most interesting ones are
currently stemming from the fields of Machine Learning, which have
already shown its relevance as the basis for MDSS and as tools to improve
hospital inpatient care.
⢠ML comes âwith the guaranteeâ of previous success in bioinformatics,
genetics and genomics, clinical applications, medical decision support and
clinical diagnosis, oncology, psychiatry and neurological disorders, or
cytopathology, to name a few diverse areas.
⢠Strides in ML for Critical Care have been made in different subfields,
including alarm algorithms, neonatal critical care, or sepsis management at
the ICU.
10. Machine Learning & AI at the ICU
ML / ICU
⢠It has been argued that the field should consider the need to focus as much
in data-related challenges as in the development and application of
appropriate data modelling techniques, shifting part of our focus from the
data modelling stage to the data understanding and pre-processing stages.
⢠Three main challenges, namely compartmentalization, corruption and
complexity have been put forward.
⢠Compartmentalization would include problems related to data privacy and
anonymization, data integration from potentially heterogeneous databases,
and data harmonization.
⢠Corruption would involve different types of data errors, issues of data
missingness and data imprecision (usually due to a lack of matching goals in
the data acquisition and the data modelling processes).
⢠Complexity would include issues of prediction, state estimation and data
multi-modality. This latter challenge bridges the stages of data pre-
processing and modelling.
14. Mind the Interpreters
Customer is always right âŚ
⢠An example in the ICU domain. Medical experts may only accept a
parsimonious outcome from a ML method, as they require an explainable
basis for their decision making that complies with their standard
operational guidelines, often based on simple and rigid attribute scores.
Mortality prediction due to sepsis at the ICU:
Ribas, V.J., Vellido, A., Ruiz-RodrĂguez, J.C., Rello, J. (2012) Severe sepsis mortality prediction with
logistic regression over latent factors. Expert Systems with Applications, 39(2), 1937-1943.
⢠Logistic Regression + Factor Analysis: as complex as it gets in an
application context in which standard scores (SOFA, APACHE) are routinely
used.
⢠Regardless success in prediction, end-user adoption of alternative
methods should not be expected.
⢠Beware of existing methods of interpretation âŚ
15. Mind the Interpreters
A ML tour of Sepsis
⢠Sepsis is defined as life-threatening organ dysfunction caused by a
dysregulated host response to infection and Organ dysfunction can be
identified as an acute change in total SOFA score ⼠2 points consequent
to the infection.
The third international consensus definitions for sepsis and
septic shock
Singer M., Deutschman C.S., Seymour C., et al, JAMA 315(8), 801â810, 2016
⢠Sepsis WAS defined as a clinical syndrome defined by the presence of
infection and Systemic Inflammatory Response Syndrome (SIRS).
⢠So what is the improvement?
18. Mind the Interpreters
A ML tour of Sepsis
⢠Sepsis is, therefore, detected at late stages of evolution (i.e. organ
dysfunction already present). So how are patients admitted to the ICU?
⢠Q-SOFA: assessing mental status, SBP and Respiratory Rate.
⢠A test in a Spanish ICU with 354 patients with Severe Sepsis.
⢠Factor Analysis
F1: CV through SOFA and drugs F2: Haemato SOFA.
F3: Resp. SOFA and PaO2/FiO2 F4: MV and Resp. SOFA
F5: 24h SSC bundles F6: microorganism
F7: Renal SOFA and total SOFA F8: 6h SSC bundles
F9: Dysfunctional Organs F10: Hepa. SOFA
F11: CNS and dysf. Organs. F12: loci of Sepsis and polimicrobial.
F13: APACHE II and Lact. F14: total organs in dysf.
19. Mind the Interpreters
A ML tour of Sepsis
⢠What about prognosis and RoD?
⢠Different embeddings and methods have been tested.
20. Mind the Interpreters
A ML tour of Sepsis
⢠What have we learned from all of this?
⢠It is still important to consider SIRS as it plays a role in both organ
dysfunction and RoD.
⢠Specificity in assessing RoD may be improved through the combination of
clinical parameters and indicators routinely used in the ICU.
⢠Some of the ML techniques presented (RVM or the LR over FA) may be
useful in assessing prognosis and validating/testing the usability of the
new definitions of Sepsis.