Intelligence-Based Recommendation System for Critical Stroke

Management in Intensive Care Units

Luis Garc

ıa Terriza

, Jos

e L. Risco-Mart

ın

, Jos

e L. Ayala

and Gemma Reig Rosell

Department of Computer Architecture and Automation, Universidad Complutense de Madrid, 28040 Madrid, Spain

Stroke Care Unit, Hospital Universitario de La Princesa, Spain

Keywords:

Machine Learning, Genetic Algorithms, Health Recommendation System, Death Risk Prediction, Decision

Support System.

Abstract:

This work presents an integrated recommendation system capable of providing support in healthcare critical

environments such as Intensive Care Units or Stroke Care Units using Machine Learning techniques. The

system can manage several patients by reading monitoring hemodynamic data in real-time, presenting current

death risk probability, and showing recommendations that would reduce such probability and, in some cases,

avoid death. This system introduces a novel method to produce recommendations based on genetic models

and supervised machine learning. The interface is built upon a web application where clinicians can evaluate

recommendations and straightforwardly provide feedback.

1 INTRODUCTION

Stroke is one of the leading death causes in the world.

It is a cerebrovascular disease that can produce death

and long-term severe disabilities (Wang et al., 2016).

There are two stroke sub-types: hemorrhagic and is-

chaemic. Hemorrhagic strokes are caused by a ves-

sel rupture, while ischaemic strokes are provoked

by blood clots occluding brain arteries (Alexopoulos

et al., 1999). According to the World Health Orga-

nization (WHO), 6.2 million people pass away ev-

ery day by strokes, the second main cause of death

worldwide and the third prominent cause of disability

(World Health Organization, 2018).

On the other hand, Artiﬁcial Intelligence (AI)

based applications are exponentially growing nowa-

days, used in many areas such as marketing, trans-

portation, agriculture, education, medicine, etc. Re-

garding medicine and healthcare, it is stated in (Jiang

et al., 2017) that clinicians might be replaced by AI

systems at some point, although this will not happen

in the short term. However, AI-based recommenda-

tion systems are gaining trust as assistance methods

to clinicians to help them make better decisions and,

therefore, improve the possible outcome of the patient

(Asan et al., 2020).

In this work, we propose a real-time recommen-

dation system that can be used in an Intensive Care

Unit (ICU) or a stroke care unit, designed to reduce

the death risk probability and, in some cases, avoid

death. Our framework has been conceived for patients

admitted to the hospital within the ﬁrst 48 hours after

suffering a stroke attack. The AI behind our system

uses hemodynamic data to produce a death prediction

model along with a real-time recommendation system

that can predict current exitus (death) risk and recom-

mend new possible patients’ states that would not lead

to exitus. These “states” are presented as recommen-

dations to the clinicians in the form of possible hemo-

dynamic values for the patient that should potentially

reduce the exitus risk. Obviously, the clinic staff must

carefully examine and validate this new state.

To build the predictive model, the AI module is

fed with data collected from a Philips PIIC iX mon-

itoring center installed in the stroke care unit of the

Hospital Universitario La Princesa since 2017, where

around 800 patients have been successfully moni-

tored. Patients’ data have been stored in a semi-

structured SQL Database so it can be pre-processed,

modeled, and read in real-time by this recommenda-

tion system. This paper focuses on describing the rec-

ommendation system as a high-level application de-

ployed at the stroke care unit mentioned above. Intrin-

sic technical details about the predictive model deﬁ-

nition are detailed in (Garc

ıa-Terriza et al., 2021).

The paper is organized as follows. Section 2 dis-

cusses the related work. Section 3 presents the high-

level architecture and design of the recommendation

García Terriza, L., Risco-Martín, J., Ayala, J. and Roselló, G.

Intelligence-Based Recommendation System for Critical Stroke Management in Intensive Care Units.

DOI: 10.5220/0011621000003414

In Proceedings of the 16th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2023) - Volume 3: BIOINFORMATICS, pages 131-138

ISBN: 978-989-758-631-6; ISSN: 2184-4305

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

131

system. Next, Section 4 illustrates the simulations

performed to test our hypotheses and shows the re-

sults obtained for different patients. Finally, Section

5 draws some conclusions and introduces future work.

2 RELATED WORK

There is an increasing amount of research work where

diverse artiﬁcial intelligence techniques and models

are considered in critical health environments, assist-

ing clinicians in evaluating the current health status of

the patient or recommending possible treatments, tak-

ing into account the patient’s history. In the following,

we mention some relevant contributions.

(Moghadam et al., 2020) aim to predict hypo-

tension events using a Logistic Regression model in

an ICU environment. With a minimum of 5 minutes

of physiological data monitoring, the system outputs

the eventual risk of a hypo-tension crisis in the short

term. In (Nemati et al., 2018), the authors developed a

framework based on artiﬁcial intelligence to predict a

sepsis crisis event in advance from 4 to 12 hours. Us-

ing real-time monitoring from several hospitals and

a Weilbull-Cox proportional hazard model, they ob-

tained an accuracy between 56% and 72%, depend-

ing on the temporal window and the experiment. In

the same line, (Lukaszewski et al., 2008) attempted to

predict sepsis in patients who received surgery and

were admitted to ICU afterward. Using data from

daily blood analysis and PCR markers from 92 pa-

tients, they trained a neural network model, obtaining

an average accuracy of 83% in a range from 1 to 4

days before clinical diagnosis.

Closer to our research, other studies try not only to

predict critical events but also to provide or offer rec-

ommendations to clinicians. To compare and analyze

the current state of the art with our approach, we have

included Table 1, where similar works regarding rec-

ommendation systems in ICU environments are pre-

sented. The comparison has been performed taking

into account four main aspects: (i) the system is de-

ployed in ICU, (ii) the current health status of the pa-

tient is an input to the system, (iii) the system has

real-time capabilities, and ﬁnally, (iv) the technique

or algorithm used in the study.

As Table 1 shows, (Utomo et al., 2019) train

a Bayesian model using Reinforcement Learning to

predict the best treatment for the patient in the ICU.

The resulting model is used in real-time in ICU us-

ing patient monitoring data as input for the system.

(Neloy et al., 2019) perform classiﬁcations of criti-

cal patients in ICU by using association rule mining

and K-Nearest Neighbors to reduce patient mortal-

ity risk. As the system is deployed on a Cloud en-

vironment, the system is capable of being retrained

while the system is running and replacing the previ-

ous model with the new one. In (Chen et al., 2016),

the authors propose an approach capable of provid-

ing recommendations based on the Electronic Medi-

cal Record (EMR) upon the patient’s admission. Rec-

ommendations are not based on the ongoing status of

the patient nor their current health status but their his-

torical record. (Masud et al., ) recommend medica-

tions for critical care patients by using the ﬁrst 24

hours of monitoring data and comparing them with

the historical patient database. This work produces,

therefore, a one-shot recommendation. (Varatharajah

et al., 2020) propose a recommendation system based

on reinforcement learning to support the clinical man-

agement of COVID-19 patients. The system predicts

the severity of the COVID-19-infected patients and

then helps with their treatment. Finally, (Thong et al.,

2015) aim to diagnose patients’ disease from given

symptoms. They have used intuitionistic fuzzy sets

and a recommendation system to achieve such a pur-

pose, improving standalone methods of state of the

art.

Our work proposes a real-time recommendation

system capable of predicting current death risk and

providing recommendations that would reduce such

risk. We achieve that using (i) real-time data mon-

itored by the patient, (ii) a trained machine learning

model, and (iii) a genetic algorithm. In terms of ca-

pability to track and advance the health status of the

patient, our study, along with (Utomo et al., 2019),

and (Neloy et al., 2019), are the most complete sys-

tems as they are capable not only of operating in an

ICU environment but also to read patient’s monitoring

data in real time and provide recommendations about

the control variables, each one on its corresponding

area. The rest of the works presented lack this real-

time component which is fundamental to assist the

clinicians at every moment while the patient is admit-

ted to the hospital. Another relevant advantage of our

method is the minimum required time to produce rec-

ommendations or diagnoses. While (Masud et al., )

require patient monitoring for the ﬁrst 24 hours, our

work, on the contrary, needs a minimum monitoring

time of 2 minutes and a half. Finally, previous works

have not addressed the complex process of dealing

with real-time continuous hemodynamic signals that

present artifacts and produce vast data. At the same

time, our recommendation system can manage hemo-

dynamic data from hospitalized patients.

BIOINFORMATICS 2023 - 14th International Conference on Bioinformatics Models, Methods and Algorithms

132

Table 1: Deﬁnition of hemodynamic variables used. The acronyms used are: RL (Reinforcement Learning), KNN (K-Nearest

Neighbors), DM (Data Mining), NB (Naive Bayes), FL (Fuzzy Logic), RF (Random Forest), and GA (Genetic Algorithm).

Reference System in ICU Patient’s data ingestion Real-time Algorithm

Utomo et al. Yes Yes Yes RL

Neloy et al. Yes Yes Yes KNN

Chen et al. No No No (one-shot) DM + NB

Masud et al. Yes Yes (24h) No KNN

Varatharajah et al. No No No (one-shot) RL

Thong et al. No No No FL

Our Work Yes Yes Yes RF + GA

3 SYSTEM ARCHITECTURE

AND DESIGN

The recommendation system presented in this work

uses a death risk model based on machine learning

techniques that have been proved to predict patients’

exitus probability with an F-Score performance above

98%, tested within the ﬁrst 48 hours after patient’s ad-

mission (Garc

ıa-Terriza et al., 2021). It is capable of

predicting current death risks in real time for every

patient who is being monitored. Our method uses a

genetic algorithm to ﬁnd the pseudo-optimal state of

the patient’s hemodynamic system, reducing the ex-

itus risk. The genetic algorithm evaluates different

states through a patient in-silico model deﬁned with

machine learning techniques described in the follow-

ing subsections. And the last part of the recommenda-

tion system is the presentation to the clinicians, which

is done through a web interface where clinicians may

review the current patient’s status, the death risk, and

the recommendations, so the death risk is reduced.

Figure 1 depicts our system architecture. Firstly,

hemodynamic data is consumed by the recommenda-

tion system, which is connected to a SQL Database

where the monitoring center Philips PIIC iX stores

patients’ monitored data in real time. Secondly, the

recommendation system publishes a web interface

with two main purposes. On the one hand, it pro-

vides clinicians with an interface to add static vari-

ables of the patient. On the other hand, a visual in-

terface shows current risks and recommendations for

each patient admitted to the stroke care unit. These

components are detailed in the following subsections.

3.1 Model

As mentioned before, the recommendation system

uses a machine learning model to predict the patient’s

current death risk based on the patient’s hemody-

namic variables at that moment. The machine learn-

ing model has been trained from the patient’s hemo-

Figure 1: High-level system architecture.

dynamic data and static features. The set of Hemo-

dynamic variables are described in Table 2, whereas

the static features are shown in Table 3. The categori-

cal input variables were coded as numerical variables,

so the model input is suitable for machine learning

and deep learning algorithms. Hemodynamic data

may present inaccurate measurements as it is gath-

ered from sensors. Hence data is preprocessed before

using it for training/testing or production purposes.

Preprocessing is compounded of three stages: (1)

missing values ﬁlling, (2) outlier values removal, and

(3) standardization. The ﬁrst stage substitutes missing

values (null values) with the mean of that same pa-

tient and variable. The second stage replaces values

by the mean when those values are greater or equal

to four times the standard deviation. Finally, the third

Intelligence-Based Recommendation System for Critical Stroke Management in Intensive Care Units

133

preprocessing stage consists of applying the Z-Score

standardization to the input values of the model. This

last stage is especially relevant as variables compre-

hend different orders of magnitude, which might lead

to sub-optimal training.

Algorithm selection is one of the key points of the

study. One of our objectives was to ﬁnd the best type

of machine learning or deep-learning algorithm for

this problem. We aim to predict the patient’s exitus

based on the pair of hemodynamic data and static fea-

tures. These were the algorithms tested:

• Logistic Regression

• Naive Bayes

• Support Vector Machines (SVM-SVC)

• Tree-based algorithms: Decision Trees (CART),

Random Forests, and Gradient Boosting Trees

• Distance-based algorithms: K-Nearest Neigh-

bors, Dynamic Time Warping + 1-NN.

• Multi-Layer Perceptron Neural Network

• Deep Neural networks: 1D Convolutional Neural

Network (CNN 1-D), Long-Short Term Memory

(LSTM)

Out of all these algorithms, the ones that per-

formed the best were the tree-based algorithms, fol-

lowed by the Deep Neural Networks. The ﬁnal algo-

rithm chosen, and the one used in this recommenda-

tion system was the Gradient Boosting Trees, which

performed the highest F1-Score, a 98.5%. Every

training and testing procedure was done using the K-

Fold cross-validation method with K=5 to avoid over-

ﬁtting and independent results from training and test-

ing datasets.

As input data used in this work involves a partic-

ularly relevant time component, time windows while

training models became a bottom-line condition. Af-

ter training a large variety of models within different

time windows, we found the approach of training one

model for each hour since the patient’s admission as

the most feasible on both computational and perfor-

mance aspects. Thus, every model would be most ac-

curate in its speciﬁc time window.

For further details of the data preprocessing, train-

ing/testing process, and results, see (Garc

ıa-Terriza

et al., 2021).

3.2 Recommendation Algorithm

The most challenging part of the system was ﬁnding

the patient’s clinical state to minimize the death risk

probability. As described in Table 2, seven hemody-

namic variables are continuously monitored and are

candidates to be adjusted, so the patient’s exitus odds

are reduced. Hence, our system must search for valid

combinations, from a clinical point of view, of those

seven physiologically relevant variables that would

not lead to a patient’s exitus in a short-term period.

The number of combinations of seven numerical vari-

ables is not computationally feasible in near-real-time

situations, so we ﬁgured out an algorithm to search

the solutions space and ﬁnd possible values for each

patient. To this end, we have designed a genetic

algorithm-based method that offers possible combi-

nations that meet the clinical requirements.

A genetic individual is comprised of the seven

hemodynamic variables monitored from the patient

and loaded in the exitus risk model. The initial popu-

lation is created by producing random values for each

variable in every individual. Random values genera-

tion is controlled within their respective physiological

limits. For instance, it makes no sense to generate an

individual composed of cardiac frequency above 220

bpm (beats per minute) or oxygen saturation level less

than 80% or greater than 100%. These conditions are

also fulﬁlled at the gene mutation phase. The most

relevant part of a genetic algorithm is often the ﬁtness

function, which must evaluate individuals objectively

and quantify them. In our case, the ﬁtness function is

deﬁned as the combination of probabilities from sev-

eral exitus models since we have computed different

models for different temporal monitoring windows.

Altogether provide a probability that must be mini-

mized and ensure that, according to all the models,

the individual has a p probability of becoming an exi-

tus deﬁned as:

p =

∏

i=1

(1 − m

) (1)

where m

is the death probability given by the i-

th death risk model between 0.0 and 1.0 and n is the

number of models (or temporal monitoring windows)

included in the system. As the genetic model tries to

maximize the ﬁtness function, the death probabilities

are subtracted from 1.

Our genetic algorithm uses a one-point crossover

function, where a point between the seven variables

is randomly chosen for pairs of individuals so that the

left part of the ﬁrst individual is merged with the right

part of the second individual. The right part of the

ﬁrst individual is merged with the left part of the sec-

ond one, forming two children. Elitism is applied to

ensure that the best individuals are not lost between

generations by preserving a reduced group of the best

individuals after each generation.

All the individuals that have survived at the end

provide several states of the patient that, in turn, de-

ﬁne a set of acceptable solutions to our problem, i.e., a

BIOINFORMATICS 2023 - 14th International Conference on Bioinformatics Models, Methods and Algorithms

134

Table 2: Deﬁnition of hemodynamic variables used.

Variable Abbreviation Type Role Description

Rhythm Estimation RE Categorical Input A Rhythm Indicator

VE VE Numerical Input Ventricular extra systole

CF CF Numerical Input Cardiac Frequency

Breathing Frequency BF Numerical Input Respiratory Rate

Perf Perf Numerical Input Pulmonary Perfusion

SpO2 SpO2 Numerical Input Oxygen Saturation

ST-II ST Numerical Input Syst. Time Interval Index

Table 3: Deﬁnition of patient’s static variables used.

Variable Abbreviation Type Role Description

Age Age Categorical Input Ischemic/Hemorrhagic

Gender Gender Categorical Input Male/Female

Monitoring time MT Numerical Input Monitoring observation number

Type of Stroke TS Categorical Input Ischemic/Hemorrhagic

Risk Prediction RP Numerical Output Exitus probability

Figure 2: Patient features input form.

combination of seven hemodynamic values that min-

imize the death probability of a patient in near-real-

time situations. These solutions generated by the sys-

tem are the recommendations offered to the clinicians

through the web interface.

3.3 Visual Interface

To end this Section, we will describe the web interface

and the deﬁned action protocol for clinicians.

As mentioned in previous subsections, a set of

static variables is not available automatically. Thus,

they must be added to the system manually by clin-

icians every time a patient is admitted to the stroke

care unit. Figure 2 presents the web interface where

the patient’s age, gender, and stroke-subtype features

can be inserted. Age is limited to integer numbers be-

tween 20 and 100. Gender is delimited to biological

genders, male and female. The stroke type is catego-

rized into hemorrhagic, ischaemic, and stroke mim-

ics (this last one refers to events initially diagnosed

as ischaemic stroke with non-cerebrovascular causes).

In Figure 2, the number “9858406” identiﬁes the pa-

tient’s MRN (Medical Record Number), which is au-

tomatically provided by the registration system.

Lastly, we have deﬁned an action protocol for clin-

icians comprised of a set of repeating steps for every

patient. After an acute stroke, the patient is admitted

to the stroke care unit, where they are signed up in the

inﬁrmary and begin to be monitored. The patient will

start showing up on the recommendation system web

interface, where the patient features must be inserted.

At this point, the clinicians must intervene and add the

patient’s features. Since this process is manual, the

clinicians must check the patient’s dashboard at least

once per day. Once features have been inserted, the

system will start providing the current risk and rec-

ommendations. Finally, when recommendations are

available, the clinician can analyze the suitability of

these recommendations from a medical point of view

and provide positive feedback. This feedback will be

used in the future to ﬁne-tune the recommendation

systems and improve the clinician’s satisfaction.

4 EXPERIMENTS

Our experimental work has processed patients with

acute stroke monitored at the stroke unit of the Hospi-

tal Universitario de La Princesa (Madrid, Spain). In-

clusion criteria were: all patients with acute stroke

admitted to the stroke unit of the hospital and who

are susceptible to noninvasive multiparameter moni-

toring, according to the clinical protocols in force in

the aforementioned unit. Exclusion criteria were: pa-

tients admitted to the stroke unit for scheduled proce-

dures or other processes other than the acute phase of

stroke; no possibility of monitoring (technical prob-

Intelligence-Based Recommendation System for Critical Stroke Management in Intensive Care Units

135

lems, damaged monitoring equipment, patient intol-

erance/agitation); loss of data due to error in patient

coding; poor quality of data recording.

In this section, we will ﬁrst introduce experiments

from a technical and computational point of view, and

ﬁnally, we will present the system’s web interface.

Our experimental work has covered a broad spec-

trum of patient features (age, stroke type, and gender)

to validate the accuracy of the recommendation sys-

tem. However, due to space limits, we will present

cases of study aggregating by age ranges of 10 years,

combining with the two-stroke subtypes, ischaemic

and hemorrhagic. Another combination would be the

patient’s gender. Still, to keep the table size down, we

have decided to aggregate by mean, as there are no

relevant variations between genders.

All the experiments were run using the same ge-

netic model parameters and the same set of death risk

prediction models for every case. The initial popula-

tion is 20 individuals, the individual mutation proba-

bility is 20%, and the gene mutation probability, eval-

uated for each gene compounding the individual, is

10%. The cross probability between two individuals

is 50% using the one-point cross method, which is se-

lected randomly between the existing possible cross-

points. Finally, the genetic model is executed with

a maximum of 20 generations, deﬁned after several

tests as sufﬁcient to reach a stable solution. There is

elitism where the three best individuals are preserved

between generations.

Table 4 shows aggregated results from our exper-

iments for each age range and stroke sub-type (hem-

orrhagic and ischaemic). As it may be seen, the op-

timal solution found by the algorithm (i.e., the best

individual found) presents no exitus (death) risk like-

lihood in almost every case, according to the models

used in the genetic algorithm. As age rises, the ge-

netic algorithm takes more generations to ﬁnd opti-

mal solutions, which seems reasonable as the exitus

risk grows proportionally with age. Therefore solu-

tions are harder to be found if they exist. The results

also show that the stroke sub-type is a relevant feature

of the study case, as exitus risk is higher in hemor-

rhagic cases than ischaemic, making it harder for the

algorithm to ﬁnd optimal solutions. Typically, the op-

timal solution is found within the ﬁrst ﬁve generations

when the patient is younger than 70 years or suffers

an ischaemic stroke.

In terms of computation, as individuals are not

compounded of a signiﬁcant number of genes, the

population size is relatively reduced (20 individu-

als). Usually, optimal solutions are found within

the ﬁrst ﬁve generations. Simulations are computed

fast enough to provide results in a short-term period

that offers the recommendation system near real-time

characteristics. As stated before, the solutions found

by the genetic algorithm are, at last, the recommenda-

tions provided to clinicians.

A web interface displays the formerly presented

results. The interface must provide information about

the patient’s current status, clinical recommendations

to improve such status, and a mechanism to send feed-

back about the recommendations. Figure 3 depicts the

all-patients dashboard where previous requirements

are satisﬁed. First, the static feature insertion process

required for every patient on admission may be spot-

ted in (1), stating ”Insert patient information”. Re-

garding the patient risk information, in (2) and (3),

the real-time exitus likelihood and the historical accu-

mulated exitus likelihood of the patient are displayed,

respectively. The main goal of showing the histori-

cal risk is to provide a statistical marker that can be

used to compare with the current risk and evaluate

the evolution of the patient. Next, to see the recom-

mendations generated for a patient by the system, the

clinician should click on (4). A drop-down appears

depicting a table where the last observation (5) and

recommendations (6) are displayed. Aiming to reduce

the number of recommendations, we decided to limit

it to 3 unique recommendations at maximum. When

the clinician considers the recommendation suitable

for medical criteria, the clinician must click on (7),

changing the icon’s color as shown in (8). Finally, if

the patient has been monitored for less time than the

predeﬁned temporal window or the input monitoring

data are incomplete, the system shows a message in

the interface stating, ”Wait a few more seconds for

results”.

5 CONCLUSION AND FUTURE

WORK

This study has demonstrated that our recommenda-

tion system can diagnose death risk probability in

real-time for hospitalized patients and provide recom-

mendations in near real-time, which may be of as-

sistance to clinicians. Using genetic algorithms, we

have developed a model that can ﬁnd suitable solu-

tions for each patient within a short period of time.

The system’s recommendations may help clinicians

decrease the outcomes of the stroke in ICUs or stroke

care units, and, in some cases, those recommenda-

tions may help save the patient’s life. In terms of ar-

chitecture design, this system is quite ﬂexible as its

core is based on a bundled machine learning model,

which may be changed by a different model built

of almost any other underlying technology such as,

BIOINFORMATICS 2023 - 14th International Conference on Bioinformatics Models, Methods and Algorithms

136

Table 4: Experiments results.

Age Range Stroke sub-type Generations until best individual Average risk Best individual risk

[40,50) Ischaemic 2 0.29 % 0.0 %

[40,50) Hemorrhagic 3 0.41 % 0.0005 %

[50,60) Ischaemic 1 0.5 % 0.0 %

[50,60) Hemorrhagic 3 0.55 % 0.0001 %

[60,70) Ischaemic 2 0.748 % 0.0 %

[60,70) Hemorrhagic 3 0.0085 % 0.0001 %

[70,80) Ischaemic 4 0.5 % 0.04 %

[70,80) Hemorrhagic 20 36.3 % 34.1 %

[80,90) Ischaemic 3 0.32 % 0.075 %

[80,90) Hemorrhagic 20 33.8 % 26.5 %

Figure 3: Web patient dashboard view.

for instance, deep learning or reinforcement learn-

ing, without changing the overall architecture. This

could lead to enhancements in the model section in

the future, which is crucial to reduce outcomes and

mortality. Further improvements in architecture and

computation capabilities would imply implementing

the architecture on the (Public) Cloud, adding new

parts to the architecture, and decoupling some compo-

nents. A Cloud-based architecture would reduce the

maintenance load and make more accessible the auto-

mated deployment of new retrained models based on

the feedback and real-time results.

Our next step consists of deploying the recom-

mendation system in the stroke care unit of the Hos-

pital Universitario La Princesa. This installation com-

prises the connection to the production SQL Database

so that the system can read the patient’s hemodynamic

data in real-time and the web publishing within the

hospital networking systems. Furthermore, the clin-

ician feedback through the web must be stored on a

persistent device that should be easily accessible, so

reports are available. The clinician information will

be analyzed to improve the recommendation system

procedure by using clustering and segregation of the

different patient clusters.

ACKNOWLEDGEMENTS

This research has been funded by Instituto de

Salud Carlos III (RICORS-RD21/0006/0009) and co-

ﬁnanced with FEDER Funds and/or from the Euro-

pean funds of the Recovery, Transformation and Re-

silience Plan and by NextGenerationEU. This work

is also supported by Spanish Ministry of Science and

Innovation under project PID2019-110866RB-I00.

Intelligence-Based Recommendation System for Critical Stroke Management in Intensive Care Units

137

REFERENCES

Alexopoulos, E., Dounias, G. D., and Vemmos, K. (1999).

Medical diagnosis of stroke using inductive machine

learning. In Machine learning and applications, pages

20–23.

Asan, O., Bayrak, A. E., Choudhury, A., et al. (2020). Arti-

ﬁcial intelligence and human trust in healthcare: focus

on clinicians. Journal of medical Internet research,

22(6):e15154.

Chen, J. H., Podchiyska, T., and Altman, R. B. (2016).

Orderrex: clinical order decision support and out-

come predictions by data-mining electronic medical

records. Journal of the American Medical Informatics

Association, 23(2):339–348.

Garc

ıa-Terriza, L., Risco-Mart

ın, J. L., Rosell

o, G. R.,

and Ayala, J. L. (2021). Predictive and diagnosis

models of stroke from hemodynamic signal monitor-

ing. Medical & Biological Engineering & Computing,

59(6):1325–1337.

Jiang, F., Jiang, Y., Zhi, H., Dong, Y., Li, H., Ma, S., Wang,

Y., Dong, Q., Shen, H., and Wang, Y. (2017). Arti-

ﬁcial intelligence in healthcare: past, present and fu-

ture. Stroke and Vascular Neurology, 2(4):230–243.

Lukaszewski, R. A., Yates, A. M., Jackson, M. C., Swingler,

K., Scherer, J. M., Simpson, A., Sadler, P., McQuillan,

P., Titball, R. W., Brooks, T. J., et al. (2008). Presymp-

tomatic prediction of sepsis in intensive care unit pa-

tients. Clinical and Vaccine Immunology, 15(7):1089–

1094.

Masud, M. M., Cheratta, M., and Hayawi, K. Medication

recommendation for critical care patients using patient

similarity in clinical records.

Moghadam, M. C., Abad, E. M. K., Bagherzadeh, N., Ram-

singh, D., Li, G.-P., and Kain, Z. N. (2020). A

machine-learning approach to predicting hypotensive

events in icu settings. Computers in Biology and

Medicine, 118:103626.

Neloy, A. A., Shafayat Oshman, M., Islam, M., Hossain,

M., Zahir, Z. B., et al. (2019). Content-based health

recommender system for icu patient. In International

Conference on Multi-disciplinary Trends in Artiﬁcial

Intelligence, pages 229–237. Springer.

Nemati, S., Holder, A., Razmi, F., Stanley, M. D., Clif-

ford, G. D., and Buchman, T. G. (2018). An in-

terpretable machine learning model for accurate pre-

diction of sepsis in the icu. Critical care medicine,

46(4):547.

Thong, N. T. et al. (2015). Intuitionistic fuzzy recom-

mender systems: an effective tool for medical diag-

nosis. Knowledge-Based Systems, 74:133–150.

Utomo, C. P., Kurniawati, H., Li, X., and Pokharel, S.

(2019). Personalised medicine in critical care us-

ing bayesian reinforcement learning. In International

Conference on Advanced Data Mining and Applica-

tions, pages 648–657. Springer.

Varatharajah, Y., Chen, H., Trotter, A., and Iyer, R. K.

(2020). A dynamic human-in-the-loop recommender

system for evidence-based clinical staging of covid-

19. In HealthRecSys@ RecSys, pages 21–22.

Wang, H., Naghavi, M., Allen, C., Barber, R. M., Bhutta,

Z. A., Carter, A., Casey, D. C., Charlson, F. J., Chen,

A. Z., Coates, M. M., et al. (2016). Global, re-

gional, and national life expectancy, all-cause mor-

tality, and cause-speciﬁc mortality for 249 causes

of death, 1980–2015: a systematic analysis for the

global burden of disease study 2015. The lancet,

388(10053):1459–1544.

World Health Organization (2018). The top 10 causes of

death.

BIOINFORMATICS 2023 - 14th International Conference on Bioinformatics Models, Methods and Algorithms

138