Using a deterministic matching computer routine to identify hospital episodes in a Brazilian de-identified administrative database for the analysis of obstetrics hospitalisations

Abstract

Introduction
The absence of a unique patient identifier in the Brazilian hospital administrative database prevents the identification of hospital episodes with multiple hospitalisations of the same patient.

Objectives
This study aims to evaluate the information gain by using a computer routine to identify acute Obstetrics hospital episodes and its impact on assessing marks of case severity.

Methods
The data source was a de-identified Brazilian hospital administrative database from 2017 to 2020, including hospitalisations records of women of reproductive age (10 to 49 years old) for treating acute conditions (N=16,087,490). We processed this database by combining C++ and Python routines to create a hospital episodes database. From the latter, we selected obstetrics hospital episodes from 2018 to 2019 (N = 4,926,877). We compared selected characteristics of the hospital episodes according to their type (multiple vs single records per episode), testing for differences using effect size measures. We compared relative differences in case severity marks when using the hospital episode as the unit of analysis to that of isolated hospitalisations (N = 5,018,350).

Results
Compared to single-record episodes, multiple-records episodes had longer length of stay, higher amount reimbursed, and lower proportion of discharge alive. When comparing isolated hospitalisations to hospital episodes analysis, we observed an increase in all case severity indicators, especially for hospital deaths, with an increment of 13.15%. The computer routine decreased the hospital admissions with a reason for hospital discharge that did not indicate the outcome (hospital stay or inter-hospital transfer) from 2.29% to 0.73%.

Conclusions
The deterministic matching computer routine proved valuable for identifying records that refer to the same hospital episode, which improved the assessment of severe cases.

Introduction

Hospital administrative databases have been used for severe maternal morbidity (SMM) surveillance at the population level [1], following the recommendation of SMM surveillance as a complementary strategy for reducing maternal mortality [2].

The Brazilian Unified Healthcare System, also known as SUS (from the Portuguese, Sistema Único de Saúde), was created in 1988 and structured on the constitutional principles of universal coverage, comprehensive care and citizens’ participation [3]. Under SUS, hospital care is provided by public institutions at the three levels of government (federal, state, municipal) and private institutions that deliver services under contract to the public system. Although the system is in theory universal, in 2019 28.5% of the population purchased additional coverage from the supplementary private medical system, varying according to factors such as age and state of residence [4].

Hospital admissions funded by SUS are recorded using the same information system (SUS Hospital Information System, from Portuguese Sistema de Informação Hospitalar do SUS—SIH-SUS) and under the same standards, yielding a National Database. The AIH (Hospitalisation Authorization, from the Portuguese acronym for Autorização de Internação Hospitalar) database is the core SIH-SUS table. The AIH is a claim form the hospital administrative staff fills out for each discharge based on the medical record abstract [5]. Data is continually produced, with de-identified microdata consolidated and made publicly available monthly for download [6]. Considering the high hospital delivery coverage in the country and the data availability, we proposed using SIH-SUS to monitor SMM, using the WHO (World Health Organisation) definition of potentially life-threatening conditions (PLTC) [7].

The challenges for implementing the SMM surveillance using administrative data are the criteria definition and its operationalisation in the available database [1, 8]. Using the SIH-SUS poses an additional challenge, as the AIH form does not have a unique patient identifier, preventing the identification of all hospitalisations of the same patient. Analysing the AIH forms as isolated hospitalisations may underestimate markers of case severity, such as prolonged length of stay (LOS) and the occurrence of clinical complications [9]. To overcome this limitation, we developed a deterministic matching routine to identify hospital episodes in the SIH-SUS de-identified database. We used the Manitoba Centre for Health Policy (MCHP) definition of a hospital episode: “a single, continuous stay in the hospital system, irrespective of transfers between hospitals” [10]. We excluded long-term chronic care hospitalisations according to the MCHP’s definition.

This study aims to evaluate the information gain by using a computer routine to identify acute obstetrics hospital episodes and its impact on assessing marks of case severity.

Methods

Study design

We conducted a cross-sectional study using a matching deterministic computer routine to identify obstetric acute hospital episodes with admission dates from 2018 to 2019 in the Brazilian publicly funded hospital administrative database

Data Source

We used the AIH database from SIH-SUS. Although the AIH database fields changed through time, the general table structure includes patient data (e.g. sex, date of birth, postal code, city of residence); hospital data (e.g. hospital code, municipality); hospitalisation data (e.g., AIH code, type of care, admission and discharge dates, LOS, number of days in intensive care unit [ICU], amount reimbursed, principal diagnosis, up to 12 other diagnoses, principal procedure and reason for the discharge) [11]. The AIH code is the database’s primary key, identifying one, and only one, AIH record. The field type of care has two attributes: type 1, acute and type 5, long-term. The reason for the discharge has five main attributes: discharge alive, hospital stay, inter-hospital transfer, death and administrative discharge.

The SIH-SUS supports the prospective payment system adopted in Brazil to reimburse hospitals for the inpatient services rendered. The principal procedure field is coded according to the SUS Procedures, Medications, Orthoses, Prostheses and Special Materials Management System Table [12]. It defines the amount reimbursed and the expected LOS based on an estimated resource consumption for similar cases [5]. However, whenever the patient needs additional treatment (e.g. a patient admitted for an obstetric procedure requiring surgical intervention) or a longer-than-expected hospital stay due to the characteristics of the clinical presentation, the current AIH form is closed, and a new AIH form with a new code is generated. Likewise, when the patient is transferred to another hospital due to the need for a procedure not available in the original hospital or to continue the treatment in a hospital with a more complex level of care, a new AIH form with a new code is generated [13]. Multiple AIH forms with different codes are generated for the same patient in both cases. Finally, whenever it is necessary to transfer a patient to long-term care, a new AIH form is generated, receiving the same AIH code of the original event, and code 5 (long-term care) in the AIH form type of care field. One patient can have one or multiple type 5 AIH forms, depending on the time needed to stay in the hospital to complete the treatment. In the case of multiple type 5 AIH forms, the same AIH code is assigned to all of them.

We used previously downloaded and processed databases to develop a maternal health surveillance panel [14]. In summary, due to the size of the files, for each Brazilian federative unit (26 states and the federal district) and year (from 2012 to 2020), we downloaded the AIH databases using the microdatasus R package [15], running R version 4.3.1 for Linux [16], with the R Studio interface (version 2023.06.1 for Linux) [17]. Then, we selected records according to the following criteria: female sex and age between 10 and 49 years. Since the adolescent fertility rate is high in Brazil, the Ministry of Health adopts the age range 10 to 19 years in order to monitor adolescent pregnancies [18]. Next, we merged all the year’s files of a specific federative unit. Finally, we selected only AIH records for acute care (type 1). We processed each of the 27 federative unit databases, creating two binary fields to identify whether the AIH record met the diagnosis or the procedure criteria for an obstetric admission. We assigned the value 1 to a diagnosis flag field if any diagnoses were filled out with ICD-10 (International Classification of Diseases -10) O00 to O99 codes of Chapter XV (Pregnancy, Childbirth and Puerperium), and zero otherwise. Likewise, we assigned the value 1 to a procedure flag field if the principal procedure field was filled out with any obstetric procedure codes (Supplementary Appendix 1) and zero otherwise [19].

The completeness of all fields used by the deterministic matching routine or data analysis was 100%, and all of them, including date fields, presented a consistent format throughout the entire database.

Data Processing

We carried out various sequential processes to create two national study population databases: (1) the AIH Obstetric Isolated Hospitalisations database, formed by isolated AIH records with an admission date in 2018 or 2019, and (2) the AIH Obstetric Hospital Episodes database, including episodes with at least one AIH record with an admission date in 2018 or 2019. The majority of the processes were executed using the tidyverse R package [20] running the R version 4.3.1 for Linux [16], with the R Studio interface (version 2023.06.1 for Linux) [17]. Whenever that was the case, we cited the other computational tools used.

First step

We merged all the federative units’ databases to create a national type 1 AIH (acute care) database. As we aimed to analyse the AIH records in 2018 and 2019, we selected records with admission years from 2017 to 2020. We included the years 2017 and 2020 because hospital episodes could have begun in 2017 and ended in 2018. Likewise, episodes could have begun in 2019 and ended in 2020. The selection process resulted in a type 1 AIH database with 16,087,490 records representing AIH records of women of reproductive age (10 to 49 years old) for treating acute conditions.

We created a national type 5 AIH (long-term care) database using the same procedures used to create the type 1 AIH (acute care) database, but selecting type 5 AIH records. As previously mentioned, as a patient can have multiple type 5 AIH records with the same AIH code, we selected just one record, resulting in a type 5 AIH database with 45,258 records. Next, we linked the national type 1 AIH database to the type 5 AIH database, using the AIH code as the linkage key, and updated the field type of discharge, indicating a transfer to long-term chronic care when it occurred. We also grouped the original administrative discharge code with the hospital stay code, as both indicate that the patient needs to stay in the hospital to complete the treatment (Figure 1). Hence, in this analysis, we assigned the following attributes to the field reason for the hospital discharge: discharge alive, hospital stay, inter-hospital transfer, transfer to long-term care, and death in hospital.

Figure 1: Type 1 AIH Episode Database processing flowchart.

Second step

Using the type 1 AIH database generated in the first step, we ran a deterministic matching routine to identify the hospital episodes of care. Since the operation requires processing large amounts of data, sorting and iterating through it, we opted to work with a SQL database (SQLite, widely available free and open source database) [21] as the backend for the routine developed in C⁺⁺ [22] for faster execution.

Details of the routine description and the link for the source code are presented in Supplementary Appendix 2. In short, the routine first identifies all AIH records pertaining to the same episode of the same patient in the same hospital using a matching key formed by the hospital code, the woman’s municipality of residence, and the date of birth. All the AIH records that exactly agree on the matching key and have the admission date equal to or at most one day after the discharge date of the previous AIH record get the same hospital episode code. Second, whenever an inter-hospital transfer was recorded as the reason for discharge, the routine additionally searches for AIH records that exactly agreed on the woman’s postal code, municipality of residence and date of birth. All the AIH records that exactly agree on this second matching key and have the admission date equal to or at most one day after the discharge date of the previous AIH record get the same hospital episode code. The episode codes assigned in each of the two steps are different. Hence, whenever a hospital episode comprises multiple AIH records in the same hospital and one or more AIH records in a different hospital, the routine regroups them, assigning the same episode code.

The routine output CSV file has the same number of records of the type 1 AIH database (N = 16,087,490 records), including three new fields created by the routine: an episode code assigned to all AIH records of the same episode; a flag marking whether the AIH record initiates or is part of a sequence of AIH records at the same hospital, or an inter-hospital transfer; and a number that indicates the order of the record in the sequence of the AIH records that compose the episode (Figure 1).

We only assigned the same episode code to the AIH records when the difference between the admission date of the subsequent AIH and the discharge date of the previous AIH is equal to or at most one day after. Hence, the matching routine aims to identify AIH records that comprised one acute episode of care from admission to discharge, not allowing the identification of readmissions or other AIH records unrelated to the acute episode of care. As the matching key to identify inter-hospital transfers did not include the hospital code, it allows the identification of hospital transfers irrespective of the municipality where the destination hospital is located.

The final process in this step was to run a Python [23] routine with the database generated by the C⁺⁺ routine mentioned above to produce a table in which each record represents a patient hospital episode with the attributes: the episode code, which identifies each hospital episode; the patient’s age (calculated by the formula [admission date - date of birth]/ 365.25, rounded to an integer); the admission date of the first AIH record; the date of the discharge of the last AIH record; the length of stay in days (calculated by the formula discharge date - admission date, assign the value zero for hospitalisation stays less than one day); the reason for discharge in the last AIH record; the total days in ICU; the principal diagnosis at the first hospitalisation; the principal procedure at the first hospitalisation, the episode total amount reimbursed; the total number of AIH records in the episode; and the number of AIH records in the episode that met the obstetric criteria (according to diagnosis or procedure). This process resulted in an episode database in CSV format with 15,519,953 records (Figure 1).

The routine description and the link for the source code are presented in Supplementary Appendix 2.

Third step

To create the AIH Obstetric Isolated Hospitalisation database, we selected records from the type 1 AIH database (N = 16,087,490 AIH records) that met the criteria for obstetric admission (diagnosis or procedure) and had an admission date in 2018 or 2019, resulting in a database with 5,018,350 AIH records. Whereas to create the AIH Obstetric Hospital Episodes database, we selected records from the hospital episodes database created with the Python routine (N = 15,519,953 episode records) whose date of discharge in the last AIH record in the sequence was not 2017 and the date of admission in the first AIH record in the sequence was not 2020 (N = 8,173,725). We applied these criteria because we aimed to include only hospital episodes that had at least one AIH record with admission date in 2018 or 2019.

Among the 8,173,725 episode records we selected those with at least one AIH record in the episode that met the obstetric criteria, resulting in a database with 4,926,877 episode records (Figure 2).

Figure 2: AIH Obstetric Isolated Hospitalisations and AIH Obstetric Hospital Episodes Databases processing flowchart.

Each record in the AIH Obstetric Isolated Hospitalisation database refers to one isolated AIH record. In contrast, in the AIH Obstetric Hospital Episodes database, each record refers to a hospital episode, aggregating data for all AIH records that comprise the episode.

Data Analysis

To analyse the data, we used the following R (version 4.3.1 for Linux) [16] packages with the R Studio interface (version 2023.06.1 for Linux) [17]: tidyverse [20], labelled [24], gtsummary [25], effectsize [26], and DescTools [27]. We used tidyverse to process and transform data, labelled to manipulate metadata, gtsummary to create tables, effectsize and DescTools to calculate effect size measures.

First, we presented the characteristics of the hospital episodes, including the woman’s age; the number of AIH records per hospital episode; LOS; the amount reimbursed (in US$); the reason for the hospital episode discharge, informed in the last AIH record; and the five more frequent principal diagnoses and principal procedures, informed in the first episode AIH record. Next, we compared selected characteristics of the hospital episodes according to their type (multiple vs single AIH records per episode). We evaluated the differences between the two groups, estimating the effect size using the rank-biserial correlation for continuous variables and the Cramer V for categorical variables.

We also evaluated four markers of case severity, as follows: use of ICU, prolonged LOS (above three days, which is the 75th percentile of the hospital episode distribution), high reimbursement (above US$186.19, which is the 75th percentile of the hospital episode distribution), and hospital death. As we evaluated obstetric hospitalisations with various reasons for admission (principal diagnosis and procedure), we could not use specific benchmarking to define the prolonged length of stay and high reimbursement. Hence, we used the 75th percentile cutoff point to determine these outcomes. We estimated the odds ratio (OR) and 95% confidence interval (CI) to compare episodes according to their type (multiple vs single AIH records per episode), using the single-AIH record episode as the reference.

Next, we compared the interpretation of the case severity marks (use of ICU, prolonged LOS, high reimbursement, and hospital death) when using the hospital episode as the unit of analysis (the AIH Obstetric Hospital Episodes database) to that of isolated hospitalisation (the AIH Obstetric Isolated Hospitalisation database). For this analysis, we estimated the indicators in each database separately. To assess the relative difference of indicators calculated in the two databases, we used the formula: Relative difference = (the indicator in the AIH Obstetric Hospital Episodes database - the indicator in the AIH Obstetric Isolated Hospital Admissions database) /the indicator in the AIH Obstetric Isolated Hospital Admissions database) *100.

To additionally assess the information gain by using the deterministic matching computer routine, we evaluated: (1) the number of AIH records that initially did not meet the criteria of obstetrics but were identified as part of an obstetric hospital episode; (2) the computer routine ability to decrease the number of AIH records with a reason for discharge that did not indicate the final outcome (hospital stay or inter-hospital transfer).

The lack of a gold standard precluded us from evaluating measures of linkage quality for both false negative and false positive errors, so we resorted to using two different strategies to evaluate potential inconsistencies within the multiple AIH records episodes (117,253 records). First, we examined the occurrence of AIH records subsequent in time to a record indicating a death, considering that these should be the last record in the series. Second, we evaluated the differences in the postal codes among the multiple AIH records episodes. We excluded hospital episodes that only comprised inter-hospital transfers (N = 291), as the postal code was included in the matching key used to identify inter-hospital transfers. This migth indicate that the computer routine grouped AIH records that do not belong to the same woman. We use the postal code because, except for inter-hospital transfer, it was not part of the matching key. The episodes with inconsistencies were manually examined.

Results

Table 1 depicts the characteristics of the 4,926,877 obstetric hospital episodes, with at least one AIH record presenting an admission date between 2018 and 2019.

Table 1: Characteristics of obstetric hospital episodes, Brazil 2018–2019. ¹n (%) or Median (Interquartile Range - IQR) when indicated. ²Categories of the 10th revision of the International Statistical Classification of Diseases and Related Health Problems (ICD-10). ³The SUS Procedures, Medications, Orthoses, Prostheses and Special Materials Management System Table [12].

Characteristic	N = 4,926,877 1
Woman age Median (IQR)	25 (21, 31)
Number of AIH records
Multiple	117,253 (2.4%)
Single	4,809,624 (97.6%)
Length of stay (Days) Median (IQR)	2 (2, 3)
Amount reimbursed (US$) Median (IQR)	154.36 (125.49, 186.19)
Reason for episode discharge
Discharge alive	4,888,771 (99.23%)
Hospital stay	11,791 (0.24%)
Inter-hospital transfer	24,155 (0.49%)
Transfer to long-term care	12 (0.00%)
Death in hospital	2,148 (0.04%)
Five most frequent principal diagnoses 2
Single spontaneous delivery	2,127,634 (43.2%)
Single delivery by caesarean section	608,822 (12.0%)
Spontaneous abortion	180,272 (3.7%)
Premature rupture of membranes	130,105 (2.6%)
Labour and delivery complicated by fetal stress	111,183 (2.3%)
Five most frequent principal procedures 3
Vaginal delivery	1,969,312 (40.0%)
C-section	1,367,407 (27.7%)
Tretament of complications of pregnancy	433,467 (8.8%)
Post-abortion uterine curettage	327,007 (6.6%)
C-section in high-risk pregnancy	314,050 (6.4%)

The median age of hospitalised women was 25 years. Most hospital episodes had only one AIH record (97.6%), whereas for the 117,253 remaining, 93% had two. The median LOS was two days, the median amount reimbursed was US$ 154.36, and around 99% of the women were discharged alive. Single spontaneous delivery was the principal reason for the admission.

Compared to single-AIH record episodes, multiple-AIH record episodes had longer LOS, higher amount reimbursed, and lower proportion of discharge alive (Table 2). In addition, multiple- AIH record episodes were more likely associated with the use of ICU, prolonged LOS, high amount reimbursed, and hospital death (Table 3). These results explain the relative difference observed when comparing the markers of case severity calculated using the AIH Obstetric Hospital Episodes database to the one obtained using the AIH Obstetric Isolated Hospitalisation database. An increase was observed for all indicators, especially for hospital deaths, which presented an increment of 13.15%. The number of hospitalisations was overestimated by 1.82% (Table 4).

Table 2: Characteristics of obstetric hospital episodes according to the episode type (multiple vs single AIH records per episode), Brazil 2018–2019. ¹Median (Interquartile range - IQR) or n (%) when indicated. ²Continuous variables: rank-biserial correlation; categorical variables: Cramer V. ³CI 95% - Confidence Interval 95%.

	Hospital episode type
Characteristic	Multiple N = 117,253 1	Single N = 4,809,624 1	Effect size 2
			(CI 95%) 3
Age	26 (21, 32)	25 (21, 31)	0.062 (0.059, 0.065)
Length of stay (Days)	6(4, 10)	2(2, 3)	0.760 (0.759, 0.761)
Amount reimbursed (US$)	251.01 (195.37, 347.35)	153.20 (124.89, 184.67)	0.673 (0.671, 0.674)
Reason for episode discharge	n (%)	n (%)	0.042 (0.036, 0.067)
Discharge alive	114,887 (97.98%)	4,773,884 (99.26%)
Hospital Stay	1,145 (0.98%)	10,646 (0.22%)
Inter-hospital transfer	629 (0.54%)	23,526 (0.49%)
Transfer to long-term chronic care	11(0.009%)	1 (0.000%)
Death in Hospital	581 (0.496%)	1,567 (0.033%)

Table 3: Markers of cases severity of obstetrics hospital episode, according to the episode type (multiple vs single AIH records per episode), Brazil 2018–2019. ¹n (%). ²Reference category. ³Confidence Interval 95%. ⁴ > 75th percentile of the distribution of the length of stay. ⁵ > 75th percentile of the distribution of the value reimbursed (US$).

	Hospital episode type
Characteristic	Multiple 1 N = 117,253	Single 1 N = 4,809,624 [REF] 2	Odds ratio (CI 95%) 3
Use of Intensive Care Unit	7,677 (6.55%)	19,838 (0.41%)	16.91 (16.46, 17.38)
Prolonged length of stay 4	91,434 (77.98%)	662,408 (13.77%)	22.17 (21.86, 22.48)
High reimbursement 5	92,904 (79.23%)	1,138,661 (23.67%)	12.30 (12.13, 12.48)
Death in Hospital	581 (0.496%)	1567 (0.033%)	15.27 (13.9, 16.81)

Table 4: Markers of cases severity according to the type of analysis, Brazil 2018–2019. ¹Analysis without (Isolated) or with (Hospital Episode) aggregation of records related to the same hospital episode. ²n (%). ³Relative Difference: (%Episode of Care-%Isolated)/%Isolated*100. ⁴ >75th percentile of the distribution of episodes of care.

	Type of analysis 1
Characteristic	Hospital episode 2	Isolated hospitalisation 2	Relative difference 3 (%)–1.82%
	N = 4,926,877	N = 5,018,350
Use of intensive care unit	27,515 (0.56%)	27,333 (0.54%)	+3.70%
Prolonged length of stay 4	753,842 (15.30%)	728,721(14.52%)	+5.37%
High reimbursement 4	1,231,565 (25.00%)	1,193,840(23.79%)	+5.09%
Death in hospital	2,148 (0.044%)	1,892 (0.038%)	+13.15%

The deterministic matching computer routine identified additional 21,752 AIH records that did not meet the obstetrics criteria initially. The principal diagnoses in those records were urinary tract infection, site not specified (1,403/ 21,752; 6.4%), sterilisation (1,400/ 21,752; 6.4%), and essential (primary) hypertension (899/21,752; 4.1%).

The computer routine decreased AIH records with a reason for discharge that did not indicate the outcome (hospital stay or inter-hospital transfer) from 2.29% (115,038/5,018,350) in the AIH Obstetric Isolated Hospital Admissions database to 0.73% (35,946/4,926,877) in the AIH Obstetric Hospital Episodes database, a relative decrease of around 68%.

Among the multiple AIH records episodes (117,253 records) we found 30 (0.03%) with recorded deaths dated before the last AIH record. The manual examination of those episodes showed that 15 the subsequent records to the registered death record had an organ donor principal diagnose, indicating that this was not an error, leaving 15 episodes (0.01%) with unexplained inconsistencies.

Of the 116,962 hospital episodes evaluated for postal code inconsistencies, we identified differences in 6,300 (5.4%). Among them, 4,705 (74.7%) had all AIH records with a principal diagnosis of the ICD-10 - Chapter XV (Pregnancy, Childbirth and Puerperium), suggesting they are related. Among the 110,953 hospital episodes with multiple AIH records without differences in postal code, 87,691 (79.0%) had all AIH records with a principal diagnosis of the ICD-10 - Chapter XV (Pregnancy, Childbirth and Puerperium).

Discussion

This study’s findings showed that the deterministic matching computer routine allowed for the identification of obstetric hospital episodes in a de-identified administrative database. This not only improved the assessment of severe cases but also decreased AIH records without information on the outcome. These results have a direct impact on data quality, potentially enhancing the monitoring of maternal potentially life-threatening conditions and the fitting of predictive models for severe maternal morbidity and mortality.

The analysis based on hospital episodes has been used to study selected diagnoses [28], hospital records combined with ICU records [29], hip fracture surgery [30], and all causes of hospitalisation [31]. Using this approach, we observed that the number of AIH records was overestimated by 1.82% when failing to identify the AIH records that comprise a single hospital episode. This figure is lower than the ones observed by other studies that assessed adult hospitalisations [28, 29]. These studies showed that the patient number overestimation varies according to the record linkage technique used [28], the principal diagnosis [28], and the method used to define the episode of care [29]. In addition to these explanations, our study looked at obstetric admissions, which are more restricted by pregnancy and relevant time frames, which may also explain the lower overestimation observed.

In our study, only 2.4% of hospital episodes involved multiple AIH records. Although linkage errors could not be ruled out [32], a small proportion of hospital episodes with multiple AIH records was expected as we studied obstetric hospitalisations of young women (median 25 years old, percentile 75% 31 years old). However, these episodes were more severe cases, presenting higher odds of ICU use, prolonged LOS, and higher reimbursement amounts. Compared to the isolated hospitalisations analysis, the evaluation of hospital episodes showed around a 5% increase in prolonged LOS and a higher amount reimbursed, as well as a 3.7% increase in ICU use. These increments are relevant, as SMM is a rare condition [8].

A recent validation study [19] compared data from the SIH-SUS with data from medical records, showing a high identification of obstetric admissions in the SIH-SUS when using the criteria proposed by the study but underreporting the main morbidities that are part of the WHO criteria for potentially life-threatening conditions [2], mainly related to hypertensive and haemorrhagic complications. Our study showed that essential hypertension was recorded in 4.5% of hospitalisations not initially identified as obstetric, which is an example of poor recording in the AIH form of the reason for hospital admission when an ICD-10 hypertension code that is not specific to pregnancy was used, resulting in the hospitalisation not being captured as an obstetrics event. Therefore, we chose to use severity markers that did not depend on the recording of morbidity (ICU admission, length of stay, mortality) and the assessment of the reimbursement amounts due to their relevance for the management of services. Nevertheless, using SIH-SUS for SMM surveillance will benefit from improvements in data quality.

A comparison of our results with those of other national studies is limited [33–36] as, in general, these studies evaluated maternal complications rather than the need for additional care. Not every woman with severe maternal morbidity will have a hospital episode with multiple AIH records, and therefore, these numbers are not directly comparable. Furthermore, these studies differ in study design (population or hospital base, prospective or sectional), the definition of maternal morbidity cases, and data source (interview with the woman or medical record data). In any case, the frequency of severe maternal morbidity in Brazilian studies ranges from 1.5% [35] to 21.2% [33], while our frequency of hospital episodes with multiple AIH records was estimated at 2.4%, therefore being within the range of values estimated in these other studies.

Maternal death is a rare event, with studies demonstrating under-reporting due to incomplete data or classification errors [37]. Therefore, the 13.15% increase in death as the reason for discharge when adopting an episode-of-care approach is very relevant. It is known that the registry of deaths in the SIH-SUS has limitations, with an estimated registration coverage of 71.8% of deaths of women of childbearing age when compared to the Brazilian Mortality Information System. This coverage varies according to hospital characteristics and location, with higher values in hospitals located in the more developed South region, in hospitals with greater complexity, managed by the federal government, with a higher proportion of adult beds available for hospitalisations publicly funded, and in teaching hospitals. On the contrary, the coverage is lower in hospitals with emergency wards [38]. Despite these limitations, a study in Brazil that compared the national maternal mortality ratio estimated using the Brazilian Mortality database and the estimate based on the SIH-SUS found no significant differences between the two data sources, concluding that the SIH-SUS could be an important complementary source for the study of maternal mortality in the country [39]. One advantage of using SIH-SUS is the more timely death identification, as the data is made available monthly, emphasising the importance of applying strategies, such as our matching routine, that result in greater identification of deaths in SIH-SUS.

This study has potential limitations. First, we did not evaluate the routine information gain for identifying potentially life-threatening conditions (PLTC), which would demand processing a second SIH-SUS table that records all services provided to the patient [12, 13]. However, we analysed important case severity marks related to the PLTC elected by our team to conduct the SMM surveillance using SIH-SUS [7]. Second, the lack of a gold standard precluded us from evaluating measures of linkage quality [32], for both false negative and false positive errors. Nevertheless, we evaluated data inconsistencies in the multiple AIH records episodes, showing that only a small proportion presented unexplained discrepancies. Third, this analysis only includes hospitalisations publicly funded, and its results do not apply to women assisted in the supplementary private medical system. Finally, although the routines we developed are rather specific to the databases that underlie this study, the overall logic can be adapted to processing other databases that similarly have tables with a set of fields that can be used as a compound key for a deterministic linkage process. Under circumstances where more personal identifiers are available, such as the full name and social security numbers, and similar characteristics are available to the researcher, the overall results may be even more robust.

Conclusion

In conclusion, the deterministic matching computer routine proved to be a valuable tool for identifying AIH records that comprise the same hospital episode. This not only improved the assessment of severe cases but also has the potential to enhance the overall quality of maternal healthcare data. These findings can be directly applied to improve the monitoring and prediction of severe morbidity and mortality in maternal healthcare.

Acknowledgments

This work was supported by the Bill & Melinda Gates Foundation [INV-027961] and the Brazilian Ministry of Health/DECIT/CNPq [445116/2020-0]. Under the grant conditions of the Foundation, a Creative Commons Attribution 4.0 Generic License has already been assigned to the Author Accepted Manuscript version that might arise from this submission Cláudia Medina Coeli (CMC) and Kenneth Rochel de Camargo Jr (KRCJ), were partially supported by research fellowship grants from the National Council for Scientific and Technological Development- CNPq (CMC -306668/2023-8; KRCJ 306228/2021-1).

Statement on conflicts of interest

The authors have no conflicts of interest to declare.

Ethics statement

The study was conducted using databases that were de-identified and made publicly available (open access). According to The Brazilian National Health Council (CNS) ethics Resolution no 510/2016 (April 7, 2016), the research ethics committee approval is waived.

Data availability statement

The SIH-SUS databases can be directly accessed in https://datasus.saude.gov.br/transferencia-de-arquivos/. The files analysed in the manuscript can be accessed in https://doi.org/10.7303/syn64313527.

Abbreviations

AIH	from Portuguese, Autorização de Internação Hospitalar (Hospitalisation Authorization)
CI	Confidence Interval
ICD-10	International Classification of Diseases -10
ICU	Intensive Care Unit
IQR	Interquartile Range
LOS	Length of Stay
MCHP	Manitoba Centre for Health Po
PLTC	Potentially Life-Threatening Conditions
SIH-SUS	from Portuguese Sistema de Informação Hospitalar do SUS (SUS Hospital Information System)
SMM	Severe Maternal Morbidity
SUS	from Portuguese Sistema Único de Saúde (The Brazilian Unified Healthcare System).
WHO	World Health Organisation

References

1. Kuklina EV, Goodman DA. Severe Maternal or Near Miss Morbidity: Implications for Public Health Surveillance and Clinical Audit. Clinical Obstetrics & Gynecology. 2018; 61(2):307–18. 10.1097/GRF.0000000000000375

   10.1097/GRF.0000000000000375

2. World Health Organization. Evaluating the quality of care for severe pregnancy complications: the WHO near-miss approach for maternal health. Geneva: World Health Organization; 2011. [Accessed 2024 Apr 30]. Available from: https://apps.who.int/iris/bitstream/handle/10665/44692/9789241502221_eng.pdf.

3. Castro MC, Massuda A, Almeida G, Menezes-Filho NA, Andrade MV, De Souza Noronha KVM, Rocha R, Macinko J, Hone T, Tasca R, Giovanella L, Malik AM, Werneck H, Fachini LA, Atun R. Brazil’s unified health system: the first 30 years and prospects for the future. The Lancet. 2019;394(10195):345–56. 10.1016/S0140-6736(19)31243-7

   10.1016/S0140-6736(19)31243-7

4. Instituto Brasileiro de Geografia e Estatística (IBGE). Pesquisa nacional de saúde: 2019: informações sobre domicílios, acesso e utilização dos serviços de saúde: Brasil, grandes regiões e unidades da federação. Coordenação de Trabalho e Rendimento. Rio de Janeiro: IBGE. 2020. [Accessed 2024 Apr 30]. Available from: https://biblioteca.ibge.gov.br/index.php/biblioteca-catalogo?view=detalhes&id=2101748.

5. Travassos Veras CM. Equity in the use of private hospitals contracted by a compulsory insurance scheme in the city of Rio de Janeiro, Brazil, in 1986. PhD thesis, London School of Economics and Political Science. [cited 2024 Apr 30]. Available from: http://etheses.lse.ac.uk/id/eprint/2431.

6. Instituto Brasileiro de Geografia e Estatística (IBGE). Sistema de informações hospitalares do SUS (SIH/SUS). [Accessed 2024 Apr 30]. Available from: https://ces.ibge.gov.br/base-de-dados/metadados/ministerio-da-saude/sistema-de-informacoes-hospitalares-do-sus-sih-sus.

7. Domingues RMSM, Dias MAB, Saraceni V, Pinheiro RS, Paiva NS, Coeli CM. Vigilância da morbidade materna no Brasil: contribuições para o debate. Cad Saúde Pública. 2023;39(11):e00151123. 10.1590/0102-311XPT151123

   10.1590/0102-311XPT151123

8. Kuklina EV, Ewing AC, Satten GA, Callaghan WM, Goodman DA, Ferre CD, et al. Ranked severe maternal morbidity index for population-level surveillance at delivery hospitalization based on hospital discharge data. Garzon S, editor. PLoS ONE. 2023;18(11):e0294140. 10.1371/journal.pone.0294140

   10.1371/journal.pone.0294140

9. Osman M, Quail J, Hudema N, Hu N. Using SAS to Create Episodes-of-Hospitalization for Health Services Research. SAS. 2015. Paper 3281-2015. [Accessed 2024 Apr 30]. Available from: https://support.sas.com/resources/papers/proceedings15/3281-2015.pdf.

10. Manitoba Centre for Health Policy. The Manitoba Population Research Data Repository. Hospital Episode / Hospital Episodes. Glossary Definition. [Accessed 2024 Apr] 30]. Available from: http://mchp-appserv.cpe.umanitoba.ca/viewDefinition.php?printer=Y&definitionID=103925.

11. Cerqueira DRC, Alves PP, Coelho DSC, Reis MVM, Lima AS. Uma análise da base de dados do sistema de informação hospitalar entre 2001 e 2018: dicionário dinâmico, disponibilidade dos dados e aspectos metodológicos para a produção de indicadores sobre violência. Ipea; 2019. Accessed [2024 Apr 30]. Available from: https://repositorio.ipea.gov.br/bitstream/11058/9409/1/Uma_analise_da_base_de_dados_do_sistema_de_informacao_hospitalar.pdf.

12. Brasil. Ministério da Saúde. SIGTAP - Sistema de Gerenciamento da Tabela de Procedimentos, Medicamentos e OPM do SUS. Accessed [2024 Apr 30]. Available from: http://sigtap.datasus.gov.br/tabela-unificada/app/sec/inicio.jsp.

13. Brasil. Ministério da Saúde/ Secretaria de Atenção à Saúde/ Departamento de Regulação, Avaliação e Controle/Coordenação-Geral de Sistemas de Informação – 2010. Manual técnico operacional do Sistema de Informação hospitalar – orientações técnicas. Versão 01.2012. 2012. Accessed [2024 Apr 30]. Available from: https://bvsms.saude.gov.br/bvs/publicacoes/manual_tecnico_sistema_informacao_hospitalar_sus.pdf.

14. Domingues RMSM, Rodrigues AS, Dias MAB, Saraceni V, Francisco RPV, Pinheiro RS, et al. Maternal health surveillance panel: a tool for expanding epidemiological surveillance of women’s health and its determinants. Rev bras epidemiol. 2024;27:e240009. 10.1590/1980-549720240009

    10.1590/1980-549720240009

15. Saldanha RDF, Bastos RR, Barcellos C. Microdatasus: pacote para download e pré-processamento de microdados do Departamento de Informática do SUS (DATASUS). Cad Saúde Pública. 2019;35(9):e00032419. 10.1590/0102-311X00032419

    10.1590/0102-311X00032419

16. R Core Team. R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing, Vienna, Austria. 2023. https://www.R-project.org/.

17. Posit team. RStudio: Integrated Development Environment for R. Posit Software, PBC, Boston, MA. 2023. Accessed [2024 Apr 30]. Available from: http://www.posit.co/.

18. Brasil, Ministério da Saúde. Nota Tecnica no 2/2024-CACRIAD/CGACI/DGCI/SAPS/MS. Accessed [2024 Dec 02] Available from: https://www.gov.br/saude/pt-br/centrais-de-conteudo/publicacoes/notas-tecnicas/2024/nota-tecnica-no-2-2024.pdf/view.

19. Domingues RMSM, Meijinhos LS, Guillen LCT, Dias MAB, Saraceni V, Pinheiro RS, Paiva NS, Coeli CM. Estudo de validação das internações obstétricas no Sistema de Informações Hospitalares do Sistema Único de Saúde para a vigilância da morbidade materna: Brasil, 2021-2022. Revista Epidemiologia e Serviços de Saúde; 2024. In press.

20. Wickham H, Averick M, Bryan J, Chang W, McGowan L, François R, et al. Welcome to the tidyverse. J Open Source Softw. 2019;4(43):1686. 10.21105/joss.01686

    10.21105/joss.01686

21. SQLite: SQLite Home Page. Accessed [2024 Apr 30]. Available from: https://www.sqlite.org/

22. C++: GCC online documentation. Accessed [2024 Apr 30]. Available from: https://gcc.gnu.org/onlinedocs/.

23. Python: The Pyhton Language reference. Accessed [2024 Apr 30]. Available from: https://docs.python.org/3/reference/index.html

24. Larmarange J. labelled: Manipulating Labelled Data. R package. Accessed [2024 Apr 30]. Available from: https://CRAN.R-project.org/package=labelled.

25. Sjoberg DD, Whiting K, Curry M, Lavery JA, Larmarange J. Reproducible summary tables with the gtsummary package. The R Journal. 2021;13:570–80. 10.32614/RJ-2021-053

    10.32614/RJ-2021-053

26. Ben-Shachar M, Lüdecke D, Makowski D. effectsize: Estimation of Effect Size Indices and Standardized Parameters. Journal of Open Source Software. 2020; 5(56), 2815. 10.21105/joss.02815

    10.21105/joss.02815

27. Signorell A. DescTools: Tools for Descriptive Statistics. R package version 0.99.54. 2024. Accessed [2024 Apr 30]. Available from: https://CRAN.R-project.org/package=DescTools.

28. Tyndall RM, Clarke JA, Shimmins J. An automated procedure for determining patient numbers from episode of care records. Medical Informatics. 1987;12(2):137–46. 10.3109/14639238709003563

    10.3109/14639238709003563

29. Fransoo R, Yogendran M, Olafson K, Ramsey C, McGowan KL, Garland A. Constructing episodes of inpatient care: data infrastructure for population-based research. BMC Med Res Methodol. 2012 Dec;12(1):133. 10.1186/1471-2288-12-133

    10.1186/1471-2288-12-133

30. Sheehan KJ, Sobolev B, Guy P, Bohm E, Hellsten E, Sutherland JM, et al. Constructing an episode of care from acute hospitalization records for studying effects of timing of hip fracture surgery. Journal Orthopaedic Research. 2016 Feb;34(2):197–204. 10.1002/jor.22997

    10.1002/jor.22997

31. Peng M, Li B, Southern DA, Eastwood CA, Quan H. Constructing Episodes of Inpatient Care: How to Define Hospital Transfer in Hospital Administrative Health Data? Medical Care. 2017 Jan;55(1):74–8. 10.1097/MLR.0000000000000624

    10.1097/MLR.0000000000000624

32. Doidge JC, Harron KL. Reflections on modern methods: linkage error bias. International Journal of Epidemiology. 2019 Oct 21;dyz203. 10.1093/ije/dyz203

    10.1093/ije/dyz203

33. Rosendo TS, Roncalli AG, Azevedo GD. Prevalence of Maternal Morbidity and Its Association with Socioeconomic Factors: A Population-based Survey of a City in Northeastern Brazil. Rev Bras Ginecol Obstet. 2017 Nov;39(11):587-595. 10.1055/s-0037-1606246

    10.1055/s-0037-1606246

34. Moreira DDS, Gubert MB. Healthcare and sociodemographic conditions related to severe maternal morbidity in a state representative population, Federal District, Brazil: A cross-sectional study. PLoS One. 2017 Aug 3;12:e0180849. 10.1371/journal.pone.0180849

    10.1371/journal.pone.0180849

35. Moraes AP, Barreto SM, Passos VM, Golino PS, Costa JA, Vasconcelos MX. Incidence and main causes of severe maternal morbidity in São Luís, Maranhão, Brazil: a longitudinal study. Sao Paulo Med J. 2011 May;129(3):146-52. 10.1590/s1516-31802011000300005

    10.1590/s1516-31802011000300005

36. Cecatti JG, Costa ML, Haddad SM, Parpinelli MA, Souza JP, Sousa MH, Surita FG, Pinto E Silva JL, Pacagnella RC, Passini R Jr; Brazilian Network for Surveillance of Severe Maternal Morbidity study Group. Network for Surveillance of Severe Maternal Morbidity: a powerful national collaboration generating data on maternal health outcomes and care. BJOG. 2016 May;123(6):946-53. 10.1111/1471-0528.13614

    10.1111/1471-0528.13614

37. Ahmed SMA, Cresswell JA, Say L. Incompleteness and misclassification of maternal death recording: a systematic review and meta-analysis. BMC Pregnancy Childbirth. 2023;23(1):794. 10.1186/s12884-023-06077-4

    10.1186/s12884-023-06077-4

38. Marques JA, Domingues RMSM, Dias MAB, Coeli CM, Pinheiro RS, Saraceni V. Predictive factors for the registration of deaths of women of childbearing age in the Hospital Admission System (SIH/SUS), Brazil, 2012-2020. Rev Bras Epidemiol. 2024; 27::e240051. 10.1590/1980-549720240051

    10.1590/1980-549720240051

39. Ranzani OT, Marinho MDF, Bierrenbach AL. Usefulness of the Hospital Information System for maternal mortality surveillance in Brazil. Rev bras epidemiol. 2023;26:e230007. 10.1590/1980-549720230007

    10.1590/1980-549720230007