CMSAssess, RSS, London, Nov, 2004 RECONSTRUCTING COMPLEX HEALTH SERVICE DATA WITH by Gilbert MacKenzie & Xuefang Li The Centre for Medical Statistics.

CMS Assess, RSS, London, Nov, 2004

RECONSTRUCTING COMPLEX

HEALTH SERVICE DATA WITH

by

Gilbert MacKenzie & Xuefang Li

The Centre for Medical Statistics


Introduction

Increasingly there is interest in monitoring and evaluating Hospital Performance in the NHS.

This has led to the compilation of

League Tables

Classically, such data are observational and their formal statistical evaluation is by covariate or “case-mix” adjustment

But NHS hospital performance indicators may not be patient-based or directly related to patient well-being – hence complicating “case-mix” adjustment


Performance Indicators

These may include:

Occupied Bed-days

Re-admission Rates

Medial Outlier Rates

NB: Often these are Episode – rather than patient-based


Classical Data Structure

Patient

Admission(s)

Completed Consultant Episode(s)

Typically, in any period, hospital data are of the form

But Performances Indicators are typically based on the Activity of the hospital ie, on the last two above.


North Staffs Study

Aims

To compile a Cumulative Event Patient History

(CEPH) file

For All ordinary Admissions to all Hospitals in

North Staffordshire for 1997-2000

In order to provide a patient-based analysis

of performance on a year to year basis


North Staffs Study

Data

Data Source LHA NHS Warehouse Data

Episode-Based for All Admissions to all Hospitals

in NS for 1997-2000

Some 326,236 episodes over this period.

But NHS identifier missing in 25% of

episodes !!


Missing Values for Main Variables in the 325,236 Episodes

Variables

Missing number

Missing percentage

NHS number 82906 25.49% Date of birth 3 0.00% Sex 364 0.11% Postcode 1739 0.53% Marital Status 109763 33.75% Discharge date 9038 2.78% Discharge method 1303 0.40% Discharge destination 1303 0.40% Episode end date 24 0.01% Last episode indicator 80 0.02% Hospital site 164939 50.71% Ward type 158272 48.66% Primary diagnosis 2249 0.69% Operation status 255800 78.65%


Ideal Data Structure

Record 1 Pat 1, Adm 1, E1








North Staffs Study

Patient 1

Patient 2

Patient 3


Simple Patient Matching Algorithm

NHS Matching Criteria : C= (Sex, DOB, Postcode) Step 1: Define Set A as Missing ID (n=82,906 episodes)

Step 2: Define Set B as Known ID (n=242,330 episodes)

Step 3: Use C to match Set A with Set B

Step 4: Consolidate 13,114 Matches in Set B

Step 5: Call the reduced Set A set Set A* & Set B , Set B*

Step 6: Use C to to match Set A* with Set B*

Step 7: Consolidate Matches in Set B*

Step 8. Finally Use C to match A** with A**

Step 9: Allocate new NHS numbers to residual in A***


Matching Result

Overall Result

Total Missing = 82,906 (25.5%)

After 1st Match = 69, 771 (21.5%)

After 2nd & 3rd = 46, 527 (14.3%)

Accuracy

About 4%-5% are wrongly matched

Also about 7% with known NHS numbers were

really different people (Sex, DOB, Postcode).


Data Structure

Ideal Attained *









* But variable number of records per patient

Now the Target CEPH File is a Flat SPSS System File

With one record per patient


Data Structure

Target CEPH file structure

Record 1 Pat 1, E1 E2 E1 EX

Record 2 Pat 2, E1 E1 E2 E1

Record 3 Pat 3, E1 EX EX EX.

Where

1) E’s are sets of episode data

2) E1 E2 => Relates to same Admission

3) EX => a set of system missing values

4) Within patient the E records are in

chronological order.

Regular Cases by Variables System File


Defining Complex File Structures

SPSS File type GROUPED command

File Type Grouped

File='c:\my documents\oldcare\LHA20-21new.dat'

Record= #epi_id 300-301 Case=nhs_num 1-10 missing=nowarn.

Record Type 1.

Data list /V0101 12 V0201 14-24 (A) V0301 26-33 (A) V0401 35 …

Record Type 2 .

Data list /V0102 12 V0202 14-24 (A) V0302 26-33 (A) V0402 35 …

Etc to a max of 51 episode records for the NS study

End File Type.

NB Other subcommands include: Duplicate, Skip, Ordered, Case.


Data Structure

Making the CEPH file useable

Record 1 Pat 1, E1 E2 E1 EX {index structure}

Record 2 Pat 2, E1 E1 E2 E1 {index structure}

Record 3 Pat 3, E1 EX EX EX {index structure}.

.

.

Now build up useful patient-based Performance Indicator quantities using SPSS’s powerful transformation language – use Vector and Loops to store and search frequently used quantities & addresses, eg


Examples of Index Building Comment compute SUMR - the number of episodes (records) per patient

vector vdata48=V4801 to V4851.

compute #sumr=0.

loop #j =1 to 51.

if ( not(missing (Vdata48(#j )) )) #sumr= #sumr+1.

end loop.

compute sumr=#sumr.

Comment compute SUMA - the number of admissions per patient.

Vector vdata47=V4701 to V4751.

compute #suma=0.

loop #j =1 to 51.

if ( not ( missing (vdata47(#j )) ) ) #suma=vdata47(#j ) .

end loop.

compute suma=#suma.


Examples of Index Building Cont’d

Comment compute first episode address for each admission.

Vector vdata47=V4701 to V4751.

Comment Zeroise.

Do repeat i=Adm01 to Adm51.

Compute i=0.

end repeat.

Comment Declare Missing.

missing values adm01 to adm51 (0).

Vector Adm=Adm01 to Adm51.

Comment compute address.

compute Adm(1)=1.

compute #k=1.

loop #J=2 to sumr.

do if (vdata47(#j) eq vdata47(#j-1) +1).

compute #k = #k+1.

compute adm(#k)=#j.

end if.

end loop.


North Staffs Study

Results

Data Episodes = 321788,

Admissons= 284,965

Patients = 188,745

Comprehensive patient-based Index built

covering all major NHS Performance Indicators

5 Files: 1997, 1998, 1999, 2000 & 1997-2000

Descriptive analysis by Hospital types and

diagnostic category (modelling to follow)


Table 1.1 Total Numbers of Patients, Admissions and Finished Consultant Episodes by year

1997 1998 1999 2000 Total Patients 47554 48138 46194 46859 188745 Admissions 69864 70998 71206 72897 284965 Episodes 77641 80089 80549 83509 321788

North Staffs Study


North Staffs Study

Numbers of patients by trust by year

1997 1998 1999 2000

Acute Trust 43975 44396 42006 44517

Combined Trust 4392 4051 5084 3965

Total 47554 48138 46194 46859


North Staffs Study

Table 3.3 Average & Median length of stay by disease by year in the Acute Trust

20.60 16.00 20.78 16.00 21.74 17.00 23.12 18.00

12.35 8.00 11.24 8.00 12.79 9.00 12.55 9.00

7.66 6.00 8.28 6.00 7.69 6.00 7.75 6.00

21.98 14.00 20.23 14.00 21.33 12.00 22.15 13.00

9.21 6.00 8.99 6.67 9.89 6.00 9.82 6.50

4.56 2.00 4.83 2.00 5.21 2.00 5.18 2.00

Hipfracture

Heartfailure

Myocardialinfarction

Stroke

Chronicchest

Otherdiseases

Mean Median

1997

Mean Median

1998

Mean Median

1999

Mean Median

2000


North Staffs Study

Table 3.4 Average & Median length of stay by disease by year in the Combined Trust

48.03 36.00 49.38 33.00 46.41 34.50 43.12 40.00

28.63 19.00 25.53 19.50 32.51 20.50 27.85 21.00

65.17 7.50 12.33 13.00 19.41 13.00 19.75 20.50

46.57 31.00 43.93 29.00 48.67 31.00 45.11 33.00

24.09 15.00 21.30 13.00 22.56 16.00 25.50 17.00

38.57 14.50 33.14 15.00 30.23 14.33 28.06 17.00

Hipfracture

Heartfailure

Myocardialinfarction

Stroke

Chronicchest

Otherdiseases

Mean Median

1997

Mean Median

1998

Mean Median

1999

Mean Median

2000


Table 5.4 4- weeks Readmission Rates by disease by year in the Acute Trust

1997 1998 1999 2000 hip fracture 0.20% 0.41% 0.73% 0.62% heart failure 5.44% 4.38% 1.42% 3.14% myocardial

infarction 3.03% 1.81% 0.91% 1.97%

stroke 0.45% 0.68% 0.18% 1.01% Chronic chest 7.95% 10.61% 2.84% 7.03%

Other patients 12.77% 13.25% 9.83% 14.42% all patients 12.79% 13.25% 9.61% 14.45%

North Staffs Study


Some Conclusions

The Complex File Commands Mixed, Group and Nested are very useful - flexible and safe.

Need to be revised to remove Dependence on ASCII input for complex health data – too big.

Transformation language is SPSS means Database Index can be built easily.

Patient-based Performance Indicators as a standard is an exciting prospect.

Results in North Staffs suggest that health of population is declining – leading to greater utilisation with time.

CMSAssess, RSS, London, Nov, 2004 RECONSTRUCTING COMPLEX HEALTH SERVICE DATA WITH by Gilbert MacKenzie & Xuefang Li The Centre for Medical Statistics.

Documents

e1 e1 e2 e1 record

e1 e2 e1 ex record

e2 record

patient slide

record type

set set

ideal data structure

patientbased slide