Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Modelling Intensive Care Unit Length of Stay and Mortality Using Artificial Neural Networks

Hermodsson, Rita and Wallin, Amanda (2019) FMSM01 20191
Mathematical Statistics
Abstract
A large data set, consisting of 300 000 patients’ admission data to various intensive care units in Sweden, acquired from the Swedish Intensive Care Register, was used in this work. The aim was to enable accurate Length of Stay (LOS) and mortality within 30 days predictions by using feed-forward
Artificial Neural Networks (ANNs). The LOS outcome was predicted as short versus long time stays, with the threshold of 5.5 days. The ambition is that models like the one implemented in the future can act as medical decision support, quality metric, and ease the planning of resources in intensive
care units.
As a part of the modelling process, an outlier detector was implemented in order to improve the predictions. The performances of the... (More)
A large data set, consisting of 300 000 patients’ admission data to various intensive care units in Sweden, acquired from the Swedish Intensive Care Register, was used in this work. The aim was to enable accurate Length of Stay (LOS) and mortality within 30 days predictions by using feed-forward
Artificial Neural Networks (ANNs). The LOS outcome was predicted as short versus long time stays, with the threshold of 5.5 days. The ambition is that models like the one implemented in the future can act as medical decision support, quality metric, and ease the planning of resources in intensive
care units.
As a part of the modelling process, an outlier detector was implemented in order to improve the predictions. The performances of the original ANN-model, regarding the Area Under the Curve (AUC), were 0.8265 (95% CI: ±0.0016) for LOS and 0.8826 (95% CI: ±0.0004) for mortality on a test set.
The outlier detector served in proving its concept. It was possible to identify admissions that are prone to be misclassified and by excluding the outliers, a better model performance was achieved. The performance improved to an AUC of 0.8336 (95% CI: ±0.0009) for LOS and 0.8876 (95% CI: ±0.0002)
for mortality.
Benchmarking was performed by using Random Forest, Logistic Regression, and SAPS3, where ANNs outperformed all of them. Moreover, the predictions of mortality within 30 days were more accurate than the ones for LOS, in all considered models. Additionally, the predictor importance was
assessed, where the result indicated whether a patient had ventilatory support or not was the most important predictor for LOS. For mortality, a patient’s age was the most important predictor according to the performed analysis. (Less)
Popular Abstract (Swedish)
Föreställ dig att det går att förutse vistelsetid och överlevnad inom 30 dagar
för en patient som läggs in på en intensivvårdsavdelning. Dessa mått kunnat
bidra till en djupare förståelse för intensivvård. Dessutom hade dörrar öppnats
för jämförelse av vårdavdelningar sinsemellan.
I det här projektet har intagningsdata till intensivvårdsavdelningar, in-
nehållande allt ifrån information om kroppstemperatur till ålder, använts för
att prognostisera de nämnda utfallen med hjälp av maskininlärning.
Svenska intensivvårdsregistret innehåller över 300 000 patienters intagnings-
data till olika intensivvårdsavdelningar i Sverige, samt utfallen av vårdbesöken.
Utifrån denna data har olika modeller tagits fram i syfte att prognostisera... (More)
Föreställ dig att det går att förutse vistelsetid och överlevnad inom 30 dagar
för en patient som läggs in på en intensivvårdsavdelning. Dessa mått kunnat
bidra till en djupare förståelse för intensivvård. Dessutom hade dörrar öppnats
för jämförelse av vårdavdelningar sinsemellan.
I det här projektet har intagningsdata till intensivvårdsavdelningar, in-
nehållande allt ifrån information om kroppstemperatur till ålder, använts för
att prognostisera de nämnda utfallen med hjälp av maskininlärning.
Svenska intensivvårdsregistret innehåller över 300 000 patienters intagnings-
data till olika intensivvårdsavdelningar i Sverige, samt utfallen av vårdbesöken.
Utifrån denna data har olika modeller tagits fram i syfte att prognostisera vis-
telsetid och överlevnad inom 30 dagar. De bästa prognostiseringarna uppnåddes
med artificiella neurala nätverk som kombinerat förutser båda utfallen.
Vi kom fram till att modellen kunde förbättras ytterligare genom att
använda en separat modell som först sållar ut patienter med alltför kom-
plexa sjukdomstillstånd som därför sannolikt är svåra att modellera. För de
resterande patienterna blir på så vis prognostiseringen mer tillförlitlig.
Informationen i intagningsdata som visade sig vara viktigast då vistelsetid
förutsågs var huruvida patienter ges andningsstöd i sin behandling. Vid prognos-
tisering av patienters överlevnad inom 30 dagar visade sig motsvarande faktor
vara åldern.
Med en välfungerande modell så kan resursplanering bli effektivare, vilket
skulle underlätta för den redan överbelamrade vården. Tillförlitlig prognos-
tisering skulle i framtiden kunna agera som ett beslutsstöd och därmed avlasta
läkare, vilket i sin tur skulle innebära en mer jämlik och trygg vård för patienter. (Less)
Please use this url to cite or link to this publication:
author
Hermodsson, Rita and Wallin, Amanda
supervisor
organization
course
FMSM01 20191
year
type
H2 - Master's Degree (Two Years)
subject
language
English
id
8985314
date added to LUP
2019-06-18 11:51:53
date last changed
2019-06-18 11:51:53
@misc{8985314,
  abstract     = {{A large data set, consisting of 300 000 patients’ admission data to various intensive care units in Sweden, acquired from the Swedish Intensive Care Register, was used in this work. The aim was to enable accurate Length of Stay (LOS) and mortality within 30 days predictions by using feed-forward
Artificial Neural Networks (ANNs). The LOS outcome was predicted as short versus long time stays, with the threshold of 5.5 days. The ambition is that models like the one implemented in the future can act as medical decision support, quality metric, and ease the planning of resources in intensive
care units.
As a part of the modelling process, an outlier detector was implemented in order to improve the predictions. The performances of the original ANN-model, regarding the Area Under the Curve (AUC), were 0.8265 (95% CI: ±0.0016) for LOS and 0.8826 (95% CI: ±0.0004) for mortality on a test set.
The outlier detector served in proving its concept. It was possible to identify admissions that are prone to be misclassified and by excluding the outliers, a better model performance was achieved. The performance improved to an AUC of 0.8336 (95% CI: ±0.0009) for LOS and 0.8876 (95% CI: ±0.0002)
for mortality.
Benchmarking was performed by using Random Forest, Logistic Regression, and SAPS3, where ANNs outperformed all of them. Moreover, the predictions of mortality within 30 days were more accurate than the ones for LOS, in all considered models. Additionally, the predictor importance was
assessed, where the result indicated whether a patient had ventilatory support or not was the most important predictor for LOS. For mortality, a patient’s age was the most important predictor according to the performed analysis.}},
  author       = {{Hermodsson, Rita and Wallin, Amanda}},
  language     = {{eng}},
  note         = {{Student Paper}},
  title        = {{Modelling Intensive Care Unit Length of Stay and Mortality Using Artificial Neural Networks}},
  year         = {{2019}},
}