PMID- 23823288
OWN - NLM
STAT- MEDLINE
DCOM- 20140411
LR  - 20151119
IS  - 1879-8365 (Electronic)
IS  - 0926-9630 (Linking)
VI  - 188
DP  - 2013
TI  - Automated validation of patient safety clinical incident classification: macro 
      analysis.
PG  - 52-7
AB  - Patient safety is the buzz word in healthcare. Incident Information Management 
      System (IIMS) is electronic software that stores clinical mishaps narratives in 
      places where patients are treated. It is estimated that in one state alone over 
      one million electronic text documents are available in IIMS. In this paper we 
      investigate the data density available in the fields entered to notify an 
      incident and the validity of the built in classification used by clinician to 
      categories the incidents. Waikato Environment for Knowledge Analysis (WEKA) 
      software was used to test the classes. Four statistical classifier based on J48, 
      Naive Bayes (NB), Naive Bayes Multinominal (NBM) and Support Vector Machine using 
      radial basis function (SVM_RBF) algorithms were used to validate the classes. The 
      data pool was 10,000 clinical incidents drawn from 7 hospitals in one state in 
      Australia. In first part of the study 1000 clinical incidents were selected to 
      determine type and number of fields worth investigating and in the second part 
      another 5448 clinical incidents were randomly selected to validate 13 clinical 
      incident types. Result shows 74.6% of the cells were empty and only 23 fields had 
      content over 70% of the time. The percentage correctly classified classes on four 
      algorithms using categorical dataset ranged from 42 to 49%, using free-text 
      datasets from 65% to 77% and using both datasets from 72% to 79%. Kappa statistic 
      ranged from 0.36 to 0.4. for categorical data, from 0.61 to 0.74. for free-text 
      and from 0.67 to 0.77 for both datasets. Similar increases in performance in the 
      3 experiments was noted on true positive rate, precision, F-measure and area 
      under curve (AUC) of receiver operating characteristics (ROC) scores. The study 
      demonstrates only 14 of 73 fields in IIMS have data that is usable for machine 
      learning experiments. Irrespective of the type of algorithms used when all 
      datasets are used performance was better. Classifier NBM showed best performance. 
      We think the classifier can be improved further by reclassifying the most 
      confused classes and there is scope to apply text mining tool on patient safety 
      classifications.
FAU - Gupta, Jaiprakash
AU  - Gupta J
AD  - Medical Informatics Laboratory, Sydney Information Technology, Sydney University, 
      Australia.
FAU - Patrick, Jon
AU  - Patrick J
LA  - eng
PT  - Journal Article
PL  - Netherlands
TA  - Stud Health Technol Inform
JT  - Studies in health technology and informatics
JID - 9214582
MH  - Algorithms
MH  - Australia
MH  - Bayes Theorem
MH  - Data Mining
MH  - Humans
MH  - Medical Errors/*classification
MH  - *Patient Safety
MH  - Risk Management/*methods
MH  - Software
MH  - Support Vector Machine
EDAT- 2013/07/05 06:00
MHDA- 2014/04/12 06:00
CRDT- 2013/07/05 06:00
PHST- 2013/07/05 06:00 [entrez]
PHST- 2013/07/05 06:00 [pubmed]
PHST- 2014/04/12 06:00 [medline]
PST - ppublish
SO  - Stud Health Technol Inform. 2013;188:52-7.