Big Data Expo 2015 - Care IQ
- 1. Data Mining and Clinical Decision Support in
Critical Care Units:
Bridging the 2 Solitudes?
Sven Van Poucke, MD
Dept. Anesthesiology, Intensive Care, Emergency Medicine, Pain Therapy, ZOL, Genk, Belgium
Zhongheng Zhang, MD
Dept. of Critical Care Medicine, Jinhua Municipal Central Hospital, Jinhua Hospital of Zhejiang University, Zhejiang, P.R. China.
- 7. 2015, big data in healthcare:
for whom the bell tolls?
data stream 24/7 ->information, knowledge, wisdom?
+ extra information:
optimization of treatments
reduce adverse events
reduce admission rates
identification of populations
- in its infancy: -> Google baseline study
- 9. 2015, big data in healthcare:
for whom the bell tolls
ICU:
10% inpatients beds
>10% mortality
>400% cost
Preprocessed data:
highly heterogenous, multi-source, multi-scale
big data (volume, veracity,variety, velocity)
time-stamped, variable granularity and tenability
context sensitive feature extraction by fusion of
domain knowledge and data driven methods
- 11. Methods
DUA: NIH training course “Protecting Human Research Participants”
Z.Z. and S.V.P (certification number: 1132877 and 1712927)
Data extraction:
Querybuilder interfase database
by SQL with pgADmin PostgreSQL tools (version 1.12.3)
Virtual Machine
Radoop/Hive
38 tables: LABEVENTS, D_PATIENTS, COMORBIDITY_SCORES,
ICUSTAY_DETAIL.
- 14. Data
Patients (n=11944), older than 15 years
Age on ICU admission, sex, Elixhauser comorbidity score,
Type of ICU setting,
Day 1 sequential organ failure assessment (SOFA),
Simplified Acute Physiology Score (SAPS-1),
Time of ICU admission and discharge, date of death,
All measurements of platelet count during ICU stay,
Comorbidities:
hypertension, paralysis, chronic pulmonary disease, diabetes,
renal failure, acquired immunodeficiency syndrome (AIDS), coagulopathy,
obesity and liver disease
- 17. Results
ICU mortality: 11.5% (n=1378)
Survivors: significantly younger (63.9±18.4 years vs 70.3±16.2 years; p<0.001)
significantly more male patients (57.0% vs 50.7%; p<0.0001)
Mortality was higher for patients admitted on MICU and SICU, not for patients admitted on CCU and CSRU.
SAPS-1 and SOFA scores: significantly higher in the non-survivors (13.9± 4.7 vs 19.3±5.6 and 5.4±3.4 vs 9.6±4.5)
In the non-survivor group, significantly more renal failure, complicated diabetes, coagulopathy and liver disease
The prevalence of normal platelet count on ICU admission was 73.8%.
Low platelet count in 12.5%, 2.3%, 1.2%, 0.6% of the cases, respectively for grade 1, 2, 3 and 4 thrombocytopenia
High platelet count in 9.0%, 0.3%, 0.0%, 0.3% of the cases for low, mild, severe and extreme thrombocythemia
Attribute weights and forward selection by NB classification and SVM selected 10 attributes for mortality prediction
Performance of predictive analysis (NB) resulted in an accuracy of 81.30%±3.58% and an AUC of 0.91±0.02
Performance of predictive analysis(SVM) resulted in an accuracy of 50.71%±0.18% and an AUC of 0.86±0.01
- 20. Conclusions
-MIMIC-II is a valuable public database for intensive care research
-We have successfully developed an integration in RapidMiner facilitating the access to MIMIC-II
-This is a major step forward in the analysis of clinical databases using data mining technology
-Essential is the will to communicate between medical professionals and data scientists.
-Is an essential step in the development of clinical decision support and closed loop systems.
-Extensive analysis before change in medical practice and inclusion in clinical guidelines.