CQL Performance

TL;DR

For the CQL queries analyzed here, the relative performance of query evaluation in patients/s stays the same for datasets with 1 million patients compared to datasets with 100 thousand patients if the system resources are sufficient and Blaze performs equally or better for datasets with more patient history. In all cases the performance increases with rising system resources.

Systems

The following systems, with increasing resources, were used for the performance evaluation:

System	Provider	CPU	Cores	RAM	SSD
LEA25	on-prem	EPYC 7543P	4	32 GiB	3.2 TB Intel P5600 over vSAN
LEA36	on-prem	EPYC 7543P	8	64 GiB	3.2 TB Intel P5600 over vSAN
LEA47	on-prem	EPYC 7543P	16	128 GiB	3.2 TB Intel P5600 over vSAN
LEA58	on-prem	EPYC 7543P	32	256 GiB	3.2 TB Intel P5600 over vSAN
LEA79	on-prem	EPYC 9555	128	512 GiB	12.8 TB Huawei OceanDisk 300P
A5N46	on-prem	Ryzen 9900X	24	96 GiB	4 TB Samsung 990 Pro

All systems were configured according to the Production Configuration guide.

On all systems, the heap memory and the block cache were each configured to use 1/4 of the total available RAM. Consequently, the Blaze process itself uses about half of the available system memory, leaving the remainder for the file system cache.

On A5N46, the measurements were done with direct reads in the resource database (RESOURCE_DB_DIRECT_READS=true and DB_RESOURCE_STORE_KV_THREADS=64), because its RAM is too small to keep both the index and the resource contents in the page cache. In addition, the heap was fixed with -Xms24g, so that the memory left for the page cache doesn't change while the JVM grows and shrinks its heap.

Datasets

The following datasets were used:

Dataset	History	# Pat. ¹	# Res. ²	# Obs. ³	Size on SSD
100k	10 years	100 k	104 M	59 M	110 GiB
100k-fh	full	100 k	317 M	191 M	323 GiB
1M	10 years	1 M	1044 M	593 M	1045 GiB

¹ Number of Patients, ² Total Number of Resources, ³ Number of Observations

The creation of these datasets is described in the Synthea section. The disk size was measured after a full manual compaction of the database. The actual disk size can be up to 50% higher, depending on the state of the background compaction process.

Methods

Metric

The metric analyzed here are the number of patients a system can process per second. It was chosen because the CQL evaluation performance depends heavily on the number of patients available in Blaze. The datasets contain either 100 k or 1 million patients in order to represent two relevant sizes from where an interpolation or extrapolation towards the target size should be possible. The metric patients per second itself is independent from the actual number of patients and can therefore be used to compare the two population sizes analysed here.

With a given patients per second value, its always possible to calculate the to be expected CQL evaluation duration by dividing the target systems number of patients by that number. So for example, if the metric is 100 k patients/s Blaze will need 1 second if it contains 100 k patients and 5 seconds if it contains 500 k patients.

Measurement

Measurements are taken by using blazectl's evaluate-measure subcommand. That subcommand first creates both a Measure and a Library resource and evaluates the Measure afterwards. By creating resources on every run, there is a slight performance impact due to the way LSM trees used by RocksDB work. The measure evaluation returns a MeasureReport that contains the evaluation duration in an extension which ensures that external timing effects play no role.

Before each measurement Blaze is restarted. After that multiple runs are taken. The first run is used to obtain the number of hits and print the Bloom filter usage. It's duration isn't used. After this run the script waits for 10 seconds in order to give Blaze time to settle. During this time the Java just-in-time compiler and garbage collector and to some work. After that 9 runs are done were the duration for the first two runs are discarded. From the remaining 7 durations, the average and standard deviation are calculated. From the average and the total number of patients the final metric patients per second is calculated.

Systems

All measurements are taken on the same VM that has the size of LEA58. The OS is Ubuntu 22.04.4 LTS. Docker version is 26.1.4.

Simple Code Search

In this section, CQL queries for selecting patients which have observations with a certain code are analyzed. The codes were chosen to produce a wide range of hits (number of matching patients). For the 100k dataset the hits are 2 %, 60 % and 100 % of the total number of patients, for the 100k-fh dataset the hits are 2 %, 57 % and 100 % and for the 1M dataset the hits are 2.5 %, 60 % and 100 %.

The first chart shows the results for the 100k dataset. It shows that the performance raises with the system size and declines a lot with the number of patients found (hits). This decline can be explained because the Bloom filters are most effective for small number of hits.

The second bar chart shows the results for the 100k-fh dataset which differs by the 100 k dataset in that it contains a full history of patient data instead of a history capped at 10 years. Especially the number of observations is 191 M compared to only 59 M in the 100k dataset. Comparing the two bar charts, the performance is nearly identical. So for simple code search, the performance doesn't depend on the amount of patient history.

The third bar chart shows the results for the 1M dataset. For the two bigger systems LEA47 and LEA58, the relative performance measured in patients per second is identical to the performance Blaze shows at the smaller datasets with only 100 k patients. However, the same can't be said for the two smaller systems LEA25 and LEA36, where the relative performance suffers due to memory limitations of those systems.

Data

Dataset	System	Code	# Hits	Time (s)	StdDev	Pat./s
100k	LEA25	17861-6	2 k	0.06	0.003	1.540 M
100k	LEA25	8310-5	60 k	0.25	0.014	402.3 k
100k	LEA25	72514-3	100 k	0.33	0.009	300.3 k
100k	LEA36	17861-6	2 k	0.05	0.002	1.972 M
100k	LEA36	8310-5	60 k	0.12	0.005	819.7 k
100k	LEA36	72514-3	100 k	0.18	0.004	562.5 k
100k	LEA47	17861-6	2 k	0.05	0.001	1.975 M
100k	LEA47	8310-5	60 k	0.07	0.002	1.363 M
100k	LEA47	72514-3	100 k	0.10	0.001	963.7 k
100k	LEA58	17861-6	2 k	0.05	0.001	1.942 M
100k	LEA58	8310-5	60 k	0.07	0.001	1.351 M
100k	LEA58	72514-3	100 k	0.09	0.002	1.153 M
100k-fh	LEA25	788-0	2 k	0.07	0.003	1.513 M
100k-fh	LEA25	44261-6	57 k	0.24	0.007	425.2 k
100k-fh	LEA25	72514-3	100 k	0.35	0.007	282.3 k
100k-fh	LEA36	788-0	2 k	0.05	0.002	1.913 M
100k-fh	LEA36	44261-6	57 k	0.13	0.003	756.3 k
100k-fh	LEA36	72514-3	100 k	0.21	0.006	487.2 k
100k-fh	LEA47	788-0	2 k	0.05	0.002	1.981 M
100k-fh	LEA47	44261-6	57 k	0.07	0.001	1.370 M
100k-fh	LEA47	72514-3	100 k	0.10	0.001	970.7 k
100k-fh	LEA58	788-0	2 k	0.05	0.002	1.979 M
100k-fh	LEA58	44261-6	57 k	0.07	0.002	1.367 M
100k-fh	LEA58	72514-3	100 k	0.09	0.001	1.151 M
1M	LEA25	17861-6	25 k	0.64	0.005	1.564 M
1M	LEA25	8310-5	603 k	7.36	0.041	135.9 k
1M	LEA25	72514-3	998 k	11.35	0.092	88.1 k
1M	LEA36	17861-6	25 k	0.48	0.003	2.081 M
1M	LEA36	8310-5	603 k	4.74	0.038	211.1 k
1M	LEA36	72514-3	998 k	7.15	0.032	139.8 k
1M	LEA47	17861-6	25 k	0.54	0.008	1.837 M
1M	LEA47	8310-5	603 k	0.76	0.011	1.321 M
1M	LEA47	72514-3	998 k	0.89	0.009	1.120 M
1M	LEA58	17861-6	25 k	0.44	0.009	2.252 M
1M	LEA58	8310-5	603 k	0.48	0.012	2.092 M
1M	LEA58	72514-3	998 k	0.57	0.014	1.741 M
1M	LEA79	17861-6	25 k	0.23	0.003	4.290 M
1M	LEA79	8310-5	603 k	0.26	0.009	3.782 M
1M	LEA79	72514-3	998 k	0.29	0.006	3.491 M
1M	A5N46	17861-6	25 k	0.25	0.008	4.007 M
1M	A5N46	8310-5	603 k	0.29	0.003	3.395 M
1M	A5N46	72514-3	998 k	0.38	0.017	2.666 M

Example CQL Query

library "observation-17861-6"
using FHIR version '4.0.0'
include FHIRHelpers version '4.0.0'

codesystem loinc: 'http://loinc.org'

context Patient

define InInitialPopulation:
  exists [Observation: Code '17861-6' from loinc]

The CQL queries can be executed with the following commands:

cql/search.sh observation-17861-6
cql/search.sh observation-8310-5
cql/search.sh observation-72514-3
cql/search.sh observation-788-0
cql/search.sh observation-44261-6

Code and Value Search

In this section, CQL Queries for selecting patients which have observations with a certain code and value are analyzed. The values were chosen to produce a wide range of hits (number of matching patients). The hits are 10 %, 50 % and 100 %.

The first chart shows the results for the 100k dataset. It shows the number of patients a system can process per second as described above. The performance raises with system size. As with Simple Code Search, queries with a small number of hits are much faster as queries with a large number of hits due to Bloom filter optimizations.

The second chart shows the results for the 1M dataset. Here the performance of the 10 % hits query is identical to that of the the 100k dataset. That can be explained because Bloom filter don't need much memory. However, the two smaller systems LEA25 and LEA36 show a degradation in the performance of the queries with larger number of hits. Here the memory limitations show while going into the actual database storage layer.

Data

Dataset	System	Code	Value	# Hits	Time (s)	StdDev	Pat./s
100k	LEA25	29463-7	13.6 kg	10 k	0.28	0.010	356.9 k
100k	LEA25	29463-7	75.3 kg	50 k	0.73	0.017	137.7 k
100k	LEA25	29463-7	185 kg	100 k	1.07	0.029	93.7 k
100k	LEA36	29463-7	13.6 kg	10 k	0.12	0.008	842.7 k
100k	LEA36	29463-7	75.3 kg	50 k	0.33	0.011	306.0 k
100k	LEA36	29463-7	185 kg	100 k	0.46	0.007	219.4 k
100k	LEA47	29463-7	13.6 kg	10 k	0.07	0.005	1.523 M
100k	LEA47	29463-7	75.3 kg	50 k	0.15	0.004	676.3 k
100k	LEA47	29463-7	185 kg	100 k	0.23	0.007	440.1 k
100k	LEA58	29463-7	13.6 kg	10 k	0.06	0.001	1.583 M
100k	LEA58	29463-7	75.3 kg	50 k	0.12	0.005	869.4 k
100k	LEA58	29463-7	185 kg	100 k	0.16	0.003	609.1 k
1M	LEA25	29463-7	13.6 kg	99 k	1.17	0.075	852.6 k
1M	LEA25	29463-7	75.3 kg	500 k	7.75	0.019	129.0 k
1M	LEA25	29463-7	185 kg	998 k	77.87	28.633	12.8 k
1M	LEA36	29463-7	13.6 kg	99 k	1.17	0.017	856.6 k
1M	LEA36	29463-7	75.3 kg	500 k	3.19	0.052	313.6 k
1M	LEA36	29463-7	185 kg	998 k	10.83	0.084	92.3 k
1M	LEA47	29463-7	13.6 kg	99 k	0.47	0.007	2.147 M
1M	LEA47	29463-7	75.3 kg	500 k	0.73	0.004	1.366 M
1M	LEA47	29463-7	185 kg	998 k	1.13	0.025	881.3 k
1M	LEA58	29463-7	13.6 kg	99 k	0.42	0.008	2.371 M
1M	LEA58	29463-7	75.3 kg	500 k	0.50	0.020	2.005 M
1M	LEA58	29463-7	185 kg	998 k	0.65	0.036	1.545 M
1M	LEA79	29463-7	13.6 kg	99 k	0.24	0.001	4.247 M
1M	LEA79	29463-7	75.3 kg	500 k	0.23	0.003	4.306 M
1M	LEA79	29463-7	185 kg	998 k	0.28	0.017	3.559 M
1M	A5N46	29463-7	13.6 kg	99 k	0.27	0.010	3.709 M
1M	A5N46	29463-7	75.3 kg	500 k	0.30	0.003	3.302 M
1M	A5N46	29463-7	185 kg	998 k	0.39	0.011	2.565 M

CQL Query

library "observation-body-weight-50"
using FHIR version '4.0.0'
include FHIRHelpers version '4.0.0'

codesystem loinc: 'http://loinc.org'
code "body-weight": '29463-7' from loinc

context Patient

define InInitialPopulation:
  exists [Observation: "body-weight"] O where O.value < 75.3 'kg'

The CQL query is executed with the following command:

cql/search.sh observation-body-weight-10
cql/search.sh observation-body-weight-50
cql/search.sh observation-body-weight-100

Ten Code Search

In this section, CQL queries for selecting patients which have conditions with one of 10 codes are analyzed. The codes were chosen to produce both a low number and a high number of hits. For the 100k dataset the hits are 0.4 % and 95 %, for the 100k-fh dataset the hits are 2 % and 98 % and for the 1M dataset the hits are 0.4 % and 95 %.

The first chart shows the results for the 100k dataset. The performance raises with system size. As with Simple Code Search and Code Value Search, queries with a small number of hits are much faster as queries with a large number of hits due to Bloom filter optimizations.

The second chart shows the results for the 100k-fh dataset. For the 100k-fh dataset the performance of the query with small number of hits is lower because the number of hits if actually larger (2 %) as that of the 100k dataset with 0.4 %.

The third chart shows the results for the 1M dataset. As with the Code Value Search queries the performance for the query with low number of hits is the same across all system sizes, because Bloom filters don't need much memory. However, the same can't be said for the query with 95 % hits. Here the LEA25 and LEA36 systems are clearly too small.

Data

Dataset	System	# Hits	Time (s)	StdDev	Pat./s
100k	LEA25	395	0.07	0.004	1.524 M
100k	LEA25	95 k	0.32	0.014	311.8 k
100k	LEA36	395	0.06	0.002	1.677 M
100k	LEA36	95 k	0.18	0.005	568.2 k
100k	LEA47	395	0.06	0.001	1.772 M
100k	LEA47	95 k	0.10	0.001	1.005 M
100k	LEA58	395	0.06	0.001	1.653 M
100k	LEA58	95 k	0.09	0.001	1.105 M
100k-fh	LEA25	2 k	0.15	0.008	660.7 k
100k-fh	LEA25	98 k	0.39	0.019	259.2 k
100k-fh	LEA36	2 k	0.08	0.001	1.184 M
100k-fh	LEA36	98 k	0.18	0.004	553.7 k
100k-fh	LEA47	2 k	0.06	0.001	1.578 M
100k-fh	LEA47	98 k	0.11	0.002	949.3 k
100k-fh	LEA58	2 k	0.07	0.001	1.536 M
100k-fh	LEA58	98 k	0.09	0.002	1.084 M
1M	LEA25	4 k	0.54	0.004	1.845 M
1M	LEA25	954 k	9.09	0.044	110.1 k
1M	LEA36	4 k	0.57	0.011	1.766 M
1M	LEA36	954 k	5.65	0.033	177.1 k
1M	LEA47	4 k	0.41	0.007	2.435 M
1M	LEA47	954 k	0.83	0.009	1.204 M
1M	LEA58	4 k	0.41	0.006	2.420 M
1M	LEA58	954 k	0.53	0.015	1.897 M
1M	LEA79	4 k	0.23	0.003	4.367 M
1M	LEA79	954 k	0.28	0.011	3.535 M
1M	A5N46	4 k	0.23	0.003	4.291 M
1M	A5N46	954 k	0.40	0.000	2.472 M

CQL Query Frequent

library "condition-ten-frequent"
using FHIR version '4.0.0'
include FHIRHelpers version '4.0.0'

codesystem sct: 'http://snomed.info/sct'

context Patient

define InInitialPopulation:
  exists [Condition: Code '444814009' from sct] or
  exists [Condition: Code '840544004' from sct] or
  exists [Condition: Code '840539006' from sct] or
  exists [Condition: Code '386661006' from sct] or
  exists [Condition: Code '195662009' from sct] or
  exists [Condition: Code '49727002' from sct] or
  exists [Condition: Code '10509002' from sct] or
  exists [Condition: Code '72892002' from sct] or
  exists [Condition: Code '36955009' from sct] or
  exists [Condition: Code '162864005' from sct]

cql/search.sh condition-ten-frequent

CQL Query Rare

library "condition-ten-rare"
using FHIR version '4.0.0'
include FHIRHelpers version '4.0.0'

codesystem sct: 'http://snomed.info/sct'

context Patient

define InInitialPopulation:
  exists [Condition: Code '62718007' from sct] or
  exists [Condition: Code '234466008' from sct] or
  exists [Condition: Code '288959006' from sct] or
  exists [Condition: Code '47505003' from sct] or
  exists [Condition: Code '698754002' from sct] or
  exists [Condition: Code '157265008' from sct] or
  exists [Condition: Code '15802004' from sct] or
  exists [Condition: Code '14760008' from sct] or
  exists [Condition: Code '36923009' from sct] or
  exists [Condition: Code '45816000' from sct]

cql/search.sh condition-ten-rare

Ten Code Search – ValueSet

In this section, CQL queries for selecting patients which have conditions with one of 10 codes are analyzed. The same codes as in section Ten Code Search are used. However, instead of multiple retrieve expressions, a single retrieve expression with a value set is used.

Data

Dataset	System	# Hits	Time (s)	StdDev	Pat./s
1M	LEA47	4 k	0.45	0.005	2.227 M
1M	LEA47	954 k	2.53	0.007	394.9 k
1M	LEA79	4 k	0.25	0.001	4.052 M
1M	LEA79	954 k	0.41	0.004	2.416 M
1M	A5N46	4 k	0.22	0.009	4.565 M
1M	A5N46	954 k	1.40	0.002	714.6 k

CQL Query Frequent

library "condition-ten-frequent-vs"
using FHIR version '4.0.0'
include FHIRHelpers version '4.0.0'

valueset vs: 'http://fhir.org/VCL?v1=(http://snomed.info/sct)(72892002;10509002;36955009;195662009;162864005;49727002;444814009;386661006;840539006;840544004)'

context Patient

define InInitialPopulation:
  exists [Condition: vs]

cql/search.sh condition-ten-frequent-vs

CQL Query Rare

library "condition-ten-rare-vs"
using FHIR version '4.0.0'
include FHIRHelpers version '4.0.0'

valueset vs: 'http://fhir.org/VCL?v1=(http://snomed.info/sct)(62718007;234466008;288959006;47505003;698754002;157265008;15802004;14760008;36923009;45816000)'

context Patient

define InInitialPopulation:
  exists [Condition: vs]

cql/search.sh condition-ten-rare-vs

All Code Search

Data

Dataset	System	# Hits	Time (s)	StdDev	Pat./s
100k	LEA25	99 k	0.35	0.013	289.8 k
100k	LEA36	99 k	0.19	0.003	517.7 k
100k	LEA47	99 k	0.11	0.001	891.8 k
100k	LEA58	99 k	0.10	0.001	1.050 M
100k-fh	LEA25	100 k	0.37	0.018	273.0 k
100k-fh	LEA36	100 k	0.20	0.006	506.3 k
100k-fh	LEA47	100 k	0.11	0.001	870.3 k
100k-fh	LEA58	100 k	0.10	0.001	1.038 M
1M	LEA25	995 k	9.66	0.041	103.5 k
1M	LEA36	995 k	5.96	0.015	167.7 k
1M	LEA47	995 k	0.89	0.011	1.120 M
1M	LEA58	995 k	0.60	0.018	1.674 M
1M	LEA79	995 k	0.29	0.016	3.411 M
1M	A5N46	995 k	0.46	0.002	2.188 M

CQL Query

cql/search.sh condition-all

All Code Search – ValueSet

Data

Dataset	System	# Hits	Time (s)	StdDev	Pat./s
1M	LEA47	995 k	37.31	0.387	26.8 k

CQL Query

cql/search.sh condition-all

Inpatient Stress Search

Data

Dataset	System	# Hits	Time (s)	StdDev	Pat./s
100k	LEA25	2 k	0.59	0.017	169.1 k
100k	LEA36	2 k	0.31	0.009	326.7 k
100k	LEA47	2 k	0.17	0.004	580.5 k
100k	LEA58	2 k	0.15	0.005	683.1 k
100k-fh	LEA25	2 k	1.89	0.016	52.8 k
100k-fh	LEA36	2 k	1.15	0.014	86.9 k
100k-fh	LEA47	2 k	0.78	0.004	127.9 k
100k-fh	LEA58	2 k	0.54	0.006	184.4 k
1M	LEA25	16 k	3.64	0.064	275.0 k
1M	LEA36	16 k	3.90	0.041	256.7 k
1M	LEA47	16 k	1.24	0.008	805.0 k
1M	LEA58	16 k	0.68	0.018	1.478 M
1M	LEA79	16 k	0.29	0.001	3.508 M
1M	A5N46	16 k	0.62	0.006	1.613 M

CQL Query

cql/search.sh inpatient-stress

Medication

Dataset	System	# Hits	Time (s)	StdDev	Pat./s
100k	LEA25	966	0.12	0.004	806.4 k
100k	LEA25	7 k	0.12	0.006	854.3 k
100k	LEA36	966	0.06	0.003	1.684 M
100k	LEA36	7 k	0.06	0.002	1.637 M
100k	LEA47	966	0.05	0.000	1.930 M
100k	LEA47	7 k	0.05	0.001	1.900 M
100k	LEA58	966	0.05	0.001	1.889 M
100k	LEA58	7 k	0.06	0.001	1.800 M
1M	LEA25	10 k	1.39	0.005	720.1 k
1M	LEA25	66 k	1.31	0.011	765.3 k
1M	LEA36	10 k	0.69	0.003	1.440 M
1M	LEA36	66 k	0.68	0.007	1.478 M
1M	LEA47	10 k	0.47	0.002	2.145 M
1M	LEA47	66 k	0.45	0.004	2.214 M
1M	LEA58	10 k	0.48	0.003	2.085 M
1M	LEA58	66 k	0.47	0.003	2.140 M

CQL Queries

cql/search.sh medication-1
cql/search.sh medication-7

Medication Ten

Dataset	System	# Hits	Time (s)	StdDev	Pat./s
100k	LEA25	15 k	0.19	0.009	533.7 k
100k	LEA36	15 k	0.10	0.002	1.022 M
100k	LEA47	15 k	0.07	0.002	1.486 M
100k	LEA58	15 k	0.07	0.001	1.491 M
1M	LEA25	149 k	2.83	0.026	353.4 k
1M	LEA36	149 k	1.39	0.004	719.4 k
1M	LEA47	149 k	0.72	0.006	1.393 M
1M	LEA58	149 k	0.64	0.003	1.574 M

Cardiometabolic Risk

In this section, a CQL query combining multiple criteria with time dependencies is analyzed. Unlike the queries above, which test a single criterion, it selects patients with a diabetes or prediabetes diagnosis at an age between 18 and 75 that has all of the following around it:

• a BMI of at least 30 kg/m² within one year of the diagnosis,
• an HbA1c above 6.0 % within two years after the diagnosis,
• a hypertension diagnosis before or within one year after the diagnosis,
• a systolic blood pressure above 130 mmHg within one year of the diagnosis, and
• an inpatient or emergency encounter within five years after the diagnosis.

Because every criterion relates to the same diagnosis, all conditions, observations and encounters involved have to be loaded and compared to its onset for every patient that has such a diagnosis.

The measurement was done with the CQL expression cache disabled. With the cache enabled, the duration drops to 0.21 seconds on A5N46, as soon as the Bloom filter of the exists expression is built, because only the patients in the Bloom filter have to be evaluated.

Data

Dataset	System	# Hits	Time (s)	StdDev	Pat./s
1M	LEA47	1 k	196.60	54.510	5.1 k
1M	LEA79	1 k	0.84	0.019	1.194 M
1M	A5N46	1 k	47.63	0.798	21.0 k

CQL Query

library "cardiometabolic-risk"
using FHIR version '4.0.0'
include FHIRHelpers version '4.0.0'

codesystem sct: 'http://snomed.info/sct'
codesystem loinc: 'http://loinc.org'
codesystem act: 'http://terminology.hl7.org/CodeSystem/v3-ActCode'

valueset "Diabetes or Prediabetes": 'http://fhir.org/VCL?v1=(http://snomed.info/sct)(44054006;15777000)'

code "Hypertension": '59621000' from sct
code "HbA1c": '4548-4' from loinc
code "BMI": '39156-5' from loinc
code "Blood Pressure": '85354-9' from loinc
code "Systolic Blood Pressure": '8480-6' from loinc
code "Inpatient": 'IMP' from act
code "Emergency": 'EMER' from act

context Patient

define "Metabolic Diagnosis":
  [Condition: "Diabetes or Prediabetes"] D
    where AgeInYearsAt(D.onset as dateTime) between 18 and 75

define "Obesity":
  [Observation: "BMI"] O
    where O.value >= 30 'kg/m2'

define "Elevated HbA1c":
  [Observation: "HbA1c"] O
    where O.value > 6.0 '%'

define "Elevated Systolic Blood Pressure":
  [Observation: "Blood Pressure"] O
    where exists (O.component C
      where C.code ~ ToConcept("Systolic Blood Pressure")
        and C.value > 130 'mm[Hg]')

define "Hospitalization":
  [Encounter] E
    where E.class ~ "Inpatient" or E.class ~ "Emergency"

define "Index Diagnosis":
  "Metabolic Diagnosis" D
    with "Obesity" O
      such that (O.effective as dateTime) within 1 year of (D.onset as dateTime)
    with "Elevated HbA1c" A
      such that (A.effective as dateTime) 2 years or less after (D.onset as dateTime)
    with [Condition: "Hypertension"] H
      such that (H.onset as dateTime) on or before (D.onset as dateTime) + 1 year
    with "Elevated Systolic Blood Pressure" P
      such that (P.effective as dateTime) within 1 year of (D.onset as dateTime)
    with "Hospitalization" E
      such that E.period.start 5 years or less after (D.onset as dateTime)

define InInitialPopulation:
  exists "Index Diagnosis"

cql/search.sh cardiometabolic-risk

Condition Code Stratification

Data

Dataset	System	# Hits	Time (s)	StdDev	Pat./s
100k	LEA58	5.2 M	12.79	0.325	7.8 k
1M	LEA58	52.3 M	399.64	11.966	2.5 k
1M	LEA79	52.3 M	80.92	4.550	12.4 k
1M	A5N46	52.3 M	267.65	4.829	3.7 k

CQL Queries

cql/search.sh stratifier-condition-code

Laboratory Observation Code Stratification

Data

Dataset	System	# Hits	Time (s)	StdDev	Pat./s
100k	LEA58	37.8 M	280.40	3.026	0
1M	LEA79	380.5 M	1138.85	7.699	0
1M	A5N46	380.5 M	2768.19	9.310	0

CQL Queries

cql/search.sh stratifier-observation-laboratory-code