Automatic Record Hash Coding and Linkage for Epidemiological Follow-up Data Confidentiality

A protocol is proposed to allow linkage of anonymous medical information within the framework of epidemiological follow-up studies. The protocol is composed of two steps; the first concerns the irreversible transformation of identification data, using a one-way hash function which is used after spelling processing. To avoid dictionary attacks, two large random files of keys, called pads, are introduced. The second step consists in the linkage of files rendered anonymous. The weight given to each linkage field is estimated by a mixture model, the likelihood of which being maximized with the Expectation and Maximization (EM) algorithm. The performance of this method has been assessed by comparing record linkage, based on exclusive use of the automatic procedure, with a manual linkage, obtained by the Burgundy Registry of Digestive Cancers. The result of the linkage of a file of 2,847 cancers with a file of 388,614 hospitalization stays in the Dijon university hospital showed a sensitivity of 97% and a specificity of 93%.

Improving Model Robustness with Bootstrapping

Objective: Recent results published by Coste et al. in discriminant analysis with ordinal responses showed the superiority of optimal discriminating analysis for ordinal responses (ODAO) both in terms of classification and simplicity of implementation compared to classic methods (Fisher’s discrimination, logistic regression) applied to medical data (prognostics of burns) and to simulated data. Nevertheless, the solutions obtained by ODAO may be sensitive to re-sampling (i.e the estimated coefficients by ODAO may show excessive sensitivity to the training sample). This study proposes some solutions to control the fluctuations of sampling and to ensure model stability.

Methods: We used intensive computational methods and bootstrapping, at the outset of model building in order to reduce the sampling variability of estimated coefficients. Thus, the estimation of the coefficients was not based on the minimization of a classification criterion of the training sample, but on the minimization of an aggregate criterion of bootstrapped replications of a classification criterion. Five aggregate criteria were studied.

Results: The improvement in terms of robustness appeared in 30% of the test cases with moderate training sample size and 55% of those with small training sample size.

Conclusion: Simulated test cases showed that bootstrapping can help construct more robust models in difficult classification situations and small training samples which are particularly frequent.

Decision Analysis for the Assessment of a Record Linkage Procedure

Objectives: According to European legislation, we must develop computer software allowing the linkage of medical records previously rendered anonymous. Some of them, like AUTOMATCH, are used in daily practice either to gather medical files in epidemiologic studies or for clinical purpose. In the first situation, the aim is to avoid homonymous errors, and in the second one, synonymous errors. The objective of this work is to study the effect of different parameters (number of identification variables, phonetic treatments of names, direct or probabilistic linkage procedure) on the reliability of the linkage in order to determine which strategy is the best according to the purpose of the linkage.

Methods: The assessment of the Burgundy Perinatal Network requires the linking of discharge abstracts of mothers and neonates, collected in all the hospitals of the region. Those data are used to compare direct and probabilistic linkage, using different parameterization strategies.

Results: If the linkage has to be performed in real time, so that no validation of indecisions generated by probabilistic linkage is possible, probabilistic linkage using three variables without any phonetic treatment seems to be the most appropriate approach, combined with a direct linkage using four variables applied to non-conclusive links. If a validation of indecisions is possible in an epidemiological study, probabilistic linkage using five variables, with a phonetic treatment adapted to the local language has to be preferred. For medical purpose, it should be combined with a direct linkage with four or five variables.

Conclusion: This paper reveals that the time and money available to manage indecision as well as the purpose of the linkage are of paramount importance for choosing a linkage strategy.

[The REDSIAM network]

The French national health database (SNIIRAM) proved to be very useful for epidemiology, health economics, evaluation, surveillance or public health. However, it is a complex database requiring important resources and expertise for being used. The REDSIAM network has been set up for promoting the collaboration of teams working on the Sniiram. The main aim of REDSIAM is to develop and validate methods for analyzing the Sniiram database for research, surveillance, evaluation and public health purposes by sharing the knowledge and experience of specialized teams in the fields of diseases identification from the Sniiram data. The work conducted within the network is devoted to the development and the validation of algorithms using Sniiram data for identifying specific diseases. The REDSIAM governance includes the Steering Committee composed of the main organizations in charge of producing and using the Sniiram data, the Bureau and the Technical Committee. The network is organized in thematic working groups focused on specific pathological domains, and a charter defines the rules for participation in the network, the functioning of the thematic working groups, the rules for publishing and making available algorithms. The articles in this special issue of the journal present the first results of some of the thematic working groups.

Early screening for type 2 diabetes following gestational diabetes mellitus in France: hardly any impact of the 2010 guidelines

AIMS

Women who had gestational diabetes mellitus (GDM) have a high risk of type 2 diabetes mellitus (T2DM) in the years following pregnancy. Most follow-up screening studies have been conducted in limited geographical areas leading to large variability in the results. The aim of our investigation was to measure how the publication of guidelines affected early screening for T2DM after a pregnancy with GDM during the period 2007-2013, in France.

METHODS

We conducted a retrospective cohort study in a representative sample of 1/97th of the French population using data from the "National Health Insurance Inter-Regime Information System," which collects individual hospital and non-hospital data for healthcare consumption.

RESULTS

The sample included 49,080 women who gave birth in 2007-2013. In the following 3 months, only 18.49% of women with GDM had an oral glucose tolerance test or a blood glucose test in 2007. This rate had not significantly increased in 2013 (p = 0.18). The proportion of women with GDM who had the recommended glycemic follow-up at 3 months (20.30 vs. 21.58%, p = 0.19) and 6 months (32.48 vs. 37.16%, p = 0.08) was not significantly different before the guidelines (2008-2009) and after the guidelines (2012-2013). At 12 months, the difference was significant (46.77 vs. 54.05%, p = 0.009).

CONCLUSION

Postpartum screening has improved only slightly since the guidelines and remains largely insufficient, with less than 25% of women with GDM screened in the first 3 months. In the first year after delivery, less than 60% of women were screened for T2DM.

[Study of algorithms to identify schizophrenia in the SNIIRAM database conducted by the REDSIAM network]

BACKGROUND

The aim of the REDSIAM network is to foster communication between users of French medico-administrative databases and to validate and promote analysis methods suitable for the data. Within this network, the working group "Mental and behavioral disorders" took an interest in algorithms to identify adult schizophrenia in the SNIIRAM database and inventoried identification criteria for patients with schizophrenia in these databases.

METHODS

The methodology was based on interviews with nine experts in schizophrenia concerning the procedures they use to identify patients with schizophrenia disorders in databases. The interviews were based on a questionnaire and conducted by telephone.

RESULTS

The synthesis of the interviews showed that the SNIIRAM contains various tables which allow coders to identify patients suffering from schizophrenia: chronic disease status, drugs and hospitalizations. Taken separately, these criteria were not sufficient to recognize patients with schizophrenia, an algorithm should be based on all of them. Apparently, only one-third of people living with schizophrenia benefit from the longstanding disease status. Not all patients are hospitalized, and coding for diagnoses at the hospitalization, notably for short stays in medicine, surgery or obstetrics departments, is not exhaustive. As for treatment with antipsychotics, it is not specific enough as such treatments are also prescribed to patients with bipolar disorders, or even other disorders. It seems appropriate to combine these complementary criteria, while keeping in mind out-patient care (every year 80,000 patients are seen exclusively in an outpatient setting), even if these data are difficult to link with other information. Finally, the experts made three propositions for selection algorithms of patients with schizophrenia.

CONCLUSION

Patients with schizophrenia can be relatively accurately identified using SNIIRAM data. Different combinations of the selected criteria must be used depending on the objectives and they must be related to an appropriate length of time.

Postpartum hemorrhage with transfusion: Trends, near misses, risk factors and management at the scale of a perinatal network

OBJECTIVES

To analyze temporal trends and management of postpartum hemorrhage (PPH) with transfusion and its related maternal near-miss (MNM) cases between 2006 and 2014 and to study risk factors.

MATERIAL AND METHODS

This retrospective cohort study from two prospective databases included 156,047 women giving birth in all the maternity hospitals of Burgundy. We analyzed temporal trends and the distribution of PPH with transfusion, the circumstances of transfer of patients between hospitals and factors associated with PPH with transfusion. PPH with massive blood transfusion and/or non-medical treatment was defined as MNM. Statistical analysis included Chi² tests and logistic regression for multivariate analysis.

RESULTS

The overall rate of PPH with transfusion was 7.3‰ and globally increased during the study period whereas the MNM rate did not. MNM represented 37% of patients with PPH with transfusion and 71% of transferred patients, but surgical treatments were performed before transfer. Factors associated with PPH with transfusion were maternal age>35 years (odds ratio [OR]=1.3), prematurity (OR=5.0), cesarean section (OR=4.8), placenta previa (OR=22.0), twin pregnancy (OR=6.6), HELLP syndrome (OR=17.9) and severe small-for-gestational-age infants (OR=2.0). The first four were also associated with MNM.

CONCLUSION

MNM cases of PPH rates were steady in Burgundy while rates of PPH with transfusion increased moderately.

Detection of temporal clusters of health care-associated infections or colonizations with Pseudomonas aeruginosa

We investigated temporal clusters of Pseudomonas aeruginosa cases between 2005 and 2014 in 1 French university hospital, overall and by ward, using the Kulldorff method. Clusters of positive water samples were also investigated at the whole hospital level. Our results suggest that water outlets are not closely involved in the occurrence of clusters of P aeruginosa cases.

Association between Pseudomonas aeruginosa positive water samples and healthcare-associated cases: nine-year study at one university hospital

OBJECTIVE

To study the association between the results of water samples and Pseudomonas aeruginosa healthcare-associated cases in a French university hospital.

METHODS

Generalized Estimating Equations were used on complete case and imputed datasets. The spatial unit was the building and the time unit was the quarter.

RESULTS

For the period 2004-2013, 2932 water samples were studied; 17% were positive for P. aeruginosa. A higher incidence of P. aeruginosa cases was associated with a higher proportion of positive water samples (P=0.056 in complete case analysis and P=0.031 with the imputed dataset). The association was no longer observed when haematology and intensive care units were excluded, but was significant in analyses of data concerning intensive care units alone (P<0.001).

CONCLUSION

This study suggests that water outlet contamination in hospitals can lead to an increase in healthcare-associated P. aeruginosa cases in wards dealing with susceptible patients, but does not play a significant role in other wards.