1. Developing Risk of Mortality and Early Warning Score Models using Routinely Collected Data Healthy Computing Seminar Tessy Badriyah 22 May 2013

2. Developing Risk of Mortality models We used the Biochemistry and Haematology Outcome Model (BHOM), the fields are haemoglobin, white cell count, urea, serum sodium, serum potassium, creatinine, urea / creatinine. We developed risk of mortality model using Logistic Regression. Assessing a performance of using calibration (the  2 test or chi-test) and discrimination (area under ROC curve or c-index). We focus on using decision trees as the potential method and then we compare their performance with that of logistic regression. We also consider some techniques in machine learning (LR, DT, NN, NB, SVM, KNN) in order to find alternative methods to predict risk of mortality. We conducted experiments to assess the stability of the models by using a 10-cross validation method. We use t-test statistics to assess models from cross validation. We also propose a new measurement, exhaustive method, to assess the performance of the method for predicting risk of mortality.

3. New proposed measurement performance : Exhaustive Method Algorithm 3.5 : Exhaustive method to asses performance of the model 1: for A=1 to number of records do, 2: for B=1 to number of records do, 3:% If index of A not equal index of B, compare the risk and outcome 4: if (A not equal B) then 5: % If risk of A greater than risk of B 6: if (risk(A)>risk(B) AND outcome(A)=dead AND outcome(B)=alive) then 7: success=success+1 8:% If risk of A less than risk of B 9: elseif (risk(A)<risk(B) AND outcome(A)=alive AND outcome(B)=death) then 10: success=success+1 11:% If outcome of A equal outcome of B 12: elseif (outcome(A)=outcome(B) then 13: ; // don’t do anything 14: % otherwise (if not satisfied all above condition) 15: Else 16: fail=fail+1 17: End if 18: End if 19: End for 20: End for 21: Discrimination = success / (success+fail).

4. Comparison between Decision Trees and Logistic Regression and other machine learning techniques

5. Develop early warning score models If (pulse<39) then score=3 Elseif (pulse>=39 and pulse<47) then score=1 Elseif (pulse>=47 and pulse<90) then score=0 Elseif (pulse>=90 and pulse<101) then score=1 Elseif (pulse>=101) then score=2 3210123 <3939-4647-8990-100>100 A new proposed method : DTEWS Vital sign dataset (n = 198,755 observation sets). 7 fields in the dataset : heart rate, respiration rate, systolic blood pressure, body temperature, neurological status, peripheral oxygen saturation (SpO2) and inspired oxygen concentration

6. The Score PHYSIOLOGICAL PARAMETERS 3210123 Respiratory rate <1819-2021-24>25 SpO2SpO2 <8990 -9293-9495-99>100 Any Supplemental Oxygen? No Yes Temperature ( o C)<35.835.9-36.036.1-36.436.5-37.137.2-37.9>38.0 Systolic BP (mmHg)<89 90-116117-272 >273 Heart Rate<38 39-4647-8990-100>101 Conscious level Alert (A) Voice (V) Pain (P) Unresponsi ve (U) PHYSIOLOGICAL PARAMETERS 3210123 Respiratory rate<8<8 9-1112-20 21-24>25 SpO2SpO2 <9192 -9394-95>96 Any Supplemental Oxygen? No Yes Temperature ( o C)<35.0 35.1-36.036.1-38.038.1-39.0>39.1 Systolic BP (mmHg)<9091-100101-110111-249>250 Heart Rate <4041-5051-9091-110111-130>131 Conscious level Alert (A) Voice (V) Pain (P) Unresponsi ve (U) DTEWS VIEWS

7. The Performance

8. Thank you for your attention