ARTIFICIAL INTELLIGENCE FOR
PREDICTION OF SEPSIS IN VERY LOW
BIRTH WEIGHT INFANTS
Markus Leskinen MD PhD, Neonatologist
Children’s Hospital, University of Helsinki and Helsinki University
Hospital
The sepsis case in a nutshell
It’s all about
saving babieswith data and
artificial intelligence
T h e i s s u e
Sepsis is common
problem in NICUs with
severe complications.
Detection is difficult:
Unspecific and gradual
signs.
W h a t w e d i d
Used available
clinical data and
advanced analytical
methods to identify
upcoming sepsis risk.
T h e o u t c o m e
The model is able to
identify sepsis risk 24
hours before a
clinician with high and
clinically significant
accuracy.
Neonatal intensive care unit (NICU),
Children’s Hospital, Helsinki
• Tertiary hospital serving Southern Finland
with population of 1,6 million
• 17 000 deliveries per year
• 29 patient beds, 16 intensive care beds
• More than 1600 patients per year
• 120-150 VLBW infants (BW <1500g) per year
• Centricity Critical Care (GE Healthcare)
patient monitoring system from 1999
• > 12 000 patients
• > 2000 VLBW infants
NICU and Data Generation
Patient monitor
Ventilator
Infusion pumps
aEEG
monitor
iNO
delivery
system
Data Streams in NICU
• Monitoring of vital functions
• ECG 240 Hz
• 20,7 million measurements per day
• Invasive blood pressure measurement 120 Hz
• Oxygen saturation 60 Hz
• Temperature
• Respiratory rate
• Transcutaneous pO2, pCO2
• Direct connection of ventilators and other medical equipment
• Laboratory data
• Manually registered data
• Data measured by staff
• Drug prescription
• Doctors’ and nurses’ records
Centralized information system
• Data collection
• Analysis, visualization
• Storage
• Our NICU database• 2099 VLBW infants 1999-2013
• Median gestational age 28+6 weeks,
median birth weight 1100 g
Sepsis in Newbown
• Generalized infection with bacteremia
• Early onset sepsis
• <72 h of age
• Pathogens from mother, usually GBS
• Late onset sepsis, mostly VLBW infants• >72 h of age
• Usually hospital acquired
• 12% of VLBW infants develop late sepsis during NICU stay
• Sepsis is associated with high risk of mortality and long-
term neurodevelopmental sequelae
Diagnostic challenges
• Unspecific, gradual signs• feeding problems, fatigue
• tachypnea, apneic spells, tachy- or bradycardia
• No pathognomonic lab test
• CRP, late response
• White blood cell count: leukopenia, leukocytosis
• Blood glucose, metabolic acidosis
• Blood culture – gold standard
• slow, invasive, false negatives
Current management of suspected sepsis
• Blood culture and (prophylactic) antibiotic therapy for high
risk VLBW patients with signs on suspected sepsis
• duration and drug choice based on result of blood culture and clinical
situation
• Overuse of antibiotics
• disturbed intestinal microbiome
• resistence to antibiotics
• Potential delay in antibiotic therapy because of unspecific
signs
• increased morbidity and mortality
Can artificial intelligence be used for early
diagnosis of sepsis in VLBW infants?
• Predictive machine learning models are able to detect events and
abnormalities b e f o r e n o t a b l e p a t h o l o g i c a l s y m p t o m s can
be observed by conventional means.
• Our goal was to develop a computational model for p r e d i c t i n g
n e o n a t a l s e p s i s using routinely collected patient monitoring data,
laboratory results and patient record information.
• 173 VLBW infants with proven late onset sepsis
• positive blood culture and clinical diagnosis
• Control group 1702 VLBW infants without sepsis
• 106 VLBW infants with clinical suspicion of sepsis, but with
negative blood culture
• Time zero = blood culture
• Analysis of collected data 48 h prior to blood culture for
patterns that could identify sepsis with maximal accuracy
24 h prior to blood culture
• Monitor data stored as 2 min means
• Calculations using IBM Watson
• CHAID decission tree algorithm
Retrospective sepsis analysis
Measured parameters
• Monitor data
• heart rate, respiratory rate, blood pressure, oxygen saturation,
temperature, supplemental oxygen
• 2 min averages of 10 s medians
• Manual measurements
• gestational age, sex, birth weight, actual weight, diuresis
• Lab
• blood culture, blood glucose, electrolytes, CRP, blood cell count,
blood gas analyses
Derived parameters
• Variation in heart rate and temperature during last 10 min and 1 h
• Variation in respiratory rate during last 10 min
• Min ja Max temperature, pH, base excess during last 12 h
• Diuresis (ml/h/kg) during last 12 h
• Episodes of hypoxia during last 12 h
• Oxygen saturation/need for additional oxygen
• Change in mean saturation during last 3 h
• Cumulative time of hypoxia / total time of treatment
• Percent time in of hypoxia during last 3 h
• Ratio and distribution of systolic and diastolic blood pressure
Sensitivity and specificity 24 h prior to
blood culture
• At 24 h the prediction model identified blood positive
sepsis with 82% sensitivity and ja 96% specificity
• Positive predictive value 0.88
• Negative predictive value 0.94
Main parameters used by the prediction
model1. Percentage of time at low oxygen saturation / 3h
2. Arterial PO2
3. Lowest capillary pH / 12h
4. Change in mean saturation during last 3 h
5. Capillary pH
6. Oxygen saturation /need for additional oxygen
7. Capillary PO2
8. Arterial BE
9. White blood cell count
10.Capillary PCO2
Timeline of sepsis risk score
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
48 44 40 36 32 28 24 21 17 13 9 5 1
Time (h) before blood culture
Patients with sepsis
Controls
Conclusions and future development
• Our algorithm can be identify sepsis in VLBW infants 24 h
earlier than regular clinical methods
• Next step is real time analysis of risk score of sepsis in
VLBW infants
• Web-based tool for clinicians
• Can other clinical complications in NICU be detected by
machine learning?
• Necrotising enterocolitis (NEC)
• Retinopathy of prematurity (ROP)
• Intraventricular hemorrhage (IVH)
Collaborators
• HUS Lastenklinikka
• Sture Andersson
• Markus Leskinen
• IBM
• Antti Heino
• Viljami Venekoski
• Mikko Laakko
• Maija Väisänen
• Laura Sutinen
Cohort, n = 2091
Sepsis positive blood culture
n = 269
No sepsis positive blood culture
n = 1822
No positive blood culture for
candida albicans, candida
parapsilosis or yeast
n = 182
No sepsis within first 72 hours of
admission
n = 175
Clinical sepsis diagnosis
n = 182
Admission time < 180 days
n = 173Admission time < 180 days
n = 1558
No clinical sepsis diagnosis
n = 1578
SEPSIS POSITIVE TARGET GROUP REFERENCE GROUP
More than 100 records per patient
n = 1517
No positive blood culture for
candida albicans, candida
parapsilosis or yeast
n = 1569
More than 100 records per patient
n = 173