Operational challenges of building a million-patient cohort from EHRs: The COhort of DIabetic patients (CODIA) on the AP-HP EDS

Presentation in the context of a workshop on the use of health data
associated paper https://hal.science/hal-04817434

Transcript

1. Operational challenges of building a million-patient cohort from EHRs: the COhort of DIabetic patients (CODIA) on the AP-HP EDS Judith Abécassis • 16/10/2024 Atelier TIDS - PariSanté Campus
2. Codia and intercept background • Type 2 diabetes is a leading cause of premature death and handicap • population with diabetes ~10% • describe and anticipate patients medical trajectories (risk stratification) • Intercept-T2D explore the role of inflammation on the transition to diabetes cohort with different stages of the disease
3. Cohort definition and extraction legal steps Selection criteria
4. What does the EDS contain? only for inpatients !!! ~ 14 million patients 38 hospitals CDW CODIA 159 M administered drugs (structured)* 76 M medical notes (text)* 70 M reimbursements (claims, structured)* 694 M biologic results (structured)* 14 M visits* 1,3M patients * on the 2017-2024 period
5. A simplified procedure of access for APHP caregivers • EDS creation validated by the CNIL • ethical and scientific committee to approve the access ◦ level of anonymization to be specified (date shuffling or not) • Extraction and provision in a project space • overall procedure around 3 months
6. Explore the APHP EDS with the interactive tool cohort360
7. Explore the APHP EDS with the interactive tool cohort360 selection based on structured data it misses outpatients!!!
8. CODIA selection criteria Based on structured data (only for inpatients) 1. Code CIM10 diabète 2. Hba1c > 6.5% or 48mmol/mol or fasting > 7mmol/mol ou 126mg/dL 3. antidiabetic drug Based on indirect data (includes outpatients) 4. visit to a nutrition-endocrinology-diabetology department i. hard to maintain, subtleties in visit definitions and dates inherent in the data structure 5. mention of insulin, hba1c or an anti-diabetes medication in medical notes mixed criteria to try to cover everyone!
9. Conclusion: cohort definition challenges • tailored to the expected patients (in or out) • only simple criteria (no ML-based phenotyping) • (bad) surprises depending on data quality (complexity, missing data, noisy data) • cost of needing more data to refine your selection (re)define explore after 2 years, no final cohort definition, but we still move on!
10. Data particularities Quality assessment Structured vs unstructured data Feature extraction from medical notes Missing data
11. An overview of structured data • many structure information, but only for inpatients during hospitalisation • a lot of noise in the data ◦ inexact diagnosis coding (type 1 or type 2 diabetes) ◦ biology results only since 2020 ◦ progressive integration as APHP services use the new hospital software ◦ date shift in the medical notes (reflecting delayed validation by the doctors) CODIA 159 M administered drugs (structured)* 76 M medical notes (text)* 70 M reimbursements (claims, structured)* 694 M biologic results (structured)* 14 M visits* 1,3M patients * on the 2017-2024 period mixed criteria to try to cover everyone!
12. An overview of medical notes • letters • consultation or hospitalization reports • prescriptions • imaging reports Extraction of concepts rule-based or machine-learning-based (NLP), models • Biomedic (drugs, biology measures, constants) Gérardin, Remaki et al. • Charlson model (comorbidities) Petit-Jean et al.
13. Information extraction in texts Source text Entity detection Entity qualification (hypothesis / family link / negation)
14. Algorithm validation • Annotation interfaces o Colors / shortcuts • Annotation guides • Medical expertise How to annotate weight in the following sentences ? • "Weight: 76kg" • "The patient lost 5kg since last visit" • "The patient weighted 95kg 2 months ago, 97 today" Non-trivial annotation : Diabetologist / data scientist collaboration :
15. Practical application to Biomedic and Charlson models to codia • it requires a separate CSE authorization • it requires additional GPUs, with funding. 2 T4 GPUs • several weeks of running of GPUs • beta-testing in exchange for more support and evaluation of the mode (but less documentation) • for post-processing: manual construction of dictionaries • based on fuzzy matching (=slow) but it runs ! 󰠁
16. NLP is game changing for the cohort content drug extraction
17. NLP is game changing for the cohort content drug extraction
18. NLP is game changing for the cohort content diagnosis extraction Unique person_id with entities retrieved with nlp Unique person_id with diabetes retrieved by nlp vs CIM10 codes
19. Conclusion: data preparation • tedious • a lot of different information depending on the patient path (data missing not at random) • a lot of computation time, engineering skills required • but development of “commons” at the level of the EDS • problem of model sharing, evaluation, annotation for more complex alignements (ATC, NABM, CIM10) but it runs ! 󰠁
20. Perspectives of cohort usage What types of studies can we achieve with CODIA?
21. An epidemiologic study: the foot cohort amputation risk of patients hospitalized with a foot ulcer (frequent complication) up to 30% of missing data
22. Quantifying inflammation in Codia: cardiac events in diabetic patients with a monocyte measure is inflammation a prognostic marker? not causal!!!
23. A new medical discipline • particular patterns in the data ◦ missing values (not at random) ◦ incomplete patient path (other hospitals) ◦ incomplete death records (linking with the CépiDc • a lot of biases ◦ due to technical issues ◦ due to the nature of the data itself • Need to develop new methods, and new objectives • Need to have good standards for data preparation
24. key ideas 1. data is not given, but constructed (4 year man so far) 2. how can we centralize and normalize cohort construction to optimize the costs and improve the quality? 3. how can we share the data faster and stronger?
25. Thank you