People aged 65 years and over are the most frequent users of acute care hospitals in Europe, and aging trends among hospital inpatients are expected to increase dramatically in the next decades [1,2]. However, acute care hospitals often deliver substandard care to older people with complex care needs [3]. Moreover, frailty, chronic multimorbidity, disability, polypharmacy, and the resulting clinical and organizational complexity of care [4] expose older inpatients to an increased risk of hospital care-related adverse events [5-7]. A literature review not only confirmed the high incidence of adverse events among older inpatients, accounting for 5% to 60% of admissions in acute care hospitals, but also highlighted the strong association between adverse events and hospital care quality, with more than 50% of these events being deemed preventable [7]. These adverse events have important consequences for older patients, as they accelerate the aging process and lead to loss of autonomy, frequent and longer hospitalization, institutionalization, and finally death [4,6,7]. Adverse events also worsen patients’ experience with hospital care and affect their quality of life [4,8]. Finally, they weigh on health services utilization and costs [4,6].

Like most countries, Switzerland and France have initiated systemic reforms to move toward a more sustainable health care system and meet the challenges of aging and chronic multimorbidity [9]. Both governments give priority to health care quality improvement in older patients and foster the provision of better data to inform health policy, promote transparency, and improve health care efficiency [10-15]. Indeed, quality and safety indicators targeting older inpatients are essential to support and drive quality improvement, as well as inform health care stakeholders. These indicators are also of great interest to compare the performance of various health systems [16]. Some commonly used indicators based on large administrative health databases (eg, hospital discharge and insurance claims databases) could help assess and compare patient safety and health care quality across hospitals and health territories in both countries. They could also enable comparisons of Swiss and French health systems’ performance in providing high-quality, safe care to vulnerable older inpatients [12,15], which are a source of “cross-country learning” and improvement [17,18]. For example, the Patient Safety Indicators (PSIs) [19], which have been developed by the US Agency for Healthcare Research and Quality (AHRQ) [19] and adopted internationally [20,21], could screen acute hospital discharge data for potentially avoidable adverse events occurring during hospitalization. Similarly, PSIs adapted to linked administrative health data (ie, hospital discharge data linked to insurance claims data) could be of use to monitor in-hospital and postdischarge adverse events, as the latter may reflect delayed and poor quality of hospital care or premature discharge from hospital [22]. Finally, some of the Assessing Care of Vulnerable Elders-3 (ACOVE-3) quality indicators could evaluate health care processes and medication management in older inpatients based on administrative data [3,23].

Among adverse events affecting older inpatients, anticoagulant-related thromboembolic and hemorrhagic adverse events are especially frequent and serious; in fact, age is one of the strongest predictors of venous or arterial thromboembolism and bleeding during anticoagulation [24-26]. Furthermore, thromboprophylaxis is frequently suboptimal in older inpatients despite the availability of professional guidelines [27]; many studies have indeed reported recurrent prescriptions of supratherapeutic doses of anticoagulants, as well as frequent underuse and rare risk-benefit assessment of anticoagulation in this population [26-30]. Finally, the growing use of anticoagulants, especially of direct oral anticoagulants, in the geriatric inpatient population and their substantial risk of toxicity and inefficacy have become an important patient safety and public health concern worldwide [25,31]. Anticoagulant-related adverse events and the quality of anticoagulation management should therefore be routinely assessed in older inpatients receiving anticoagulants for the prevention and treatment of arterial or venous thromboembolism.

Two PSIs can be used to monitor potentially avoidable perioperative thromboembolic or bleeding events in older Swiss and French inpatients: PSI-12, Perioperative Pulmonary Embolism or Deep Vein Thrombosis Rate, and PSI-09, Perioperative Hemorrhage or Hematoma Rate [19,20,32-34]. Indeed, although both indicators target adult inpatients, most corresponding adverse events affect inpatients aged 65 years and over [20]. Moreover, PSI-12 has been adapted to coding systems of various countries, including Switzerland and France [21,35]. Lastly, both algorithms have been extended to capture both in-hospital and postdischarge adverse events based on hospital discharge data linked to outpatient claims data [22]. Besides these perioperative PSIs, others should be developed to monitor adverse events related to venous thromboembolism curative treatments and thromboprophylaxis in at-risk medical conditions such as severe acute infection or atrial fibrillation. Case-mix adjustment of PSIs should also be considered to account for differences in older inpatients’ clinical risks or disease severity at admission [4] and enable comparisons across hospitals or geographic areas [36,37]. In particular, PSIs should be adjusted for older inpatients’ risk factors for both hospital care-related adverse events (eg, frailty, chronic multimorbidity, disability, or polypharmacy) and thromboembolic or hemorrhagic adverse events (eg, age ≥75 years, renal or liver failure, inherited or acquired disorders of hemostasis, malignancy). Thus, besides the Charlson comorbidity index [38] and the updated chronic disease score [39], which have already been adapted to Swiss and French administrative health data, other comorbidity indexes [40], proxy measures of frailty and disability [41,42], and individual risk scores of thromboembolism or hemorrhage [43,44] should also be developed or adapted [45].

Regarding process quality metrics, three ACOVE-3 indicators may be used to evaluate warfarin prescription and surveillance in older patients with heart failure or atrial fibrillation [23]. Additional indicators should be developed to assess the management of other anticoagulants, including direct oral anticoagulants.

This research project aims to develop and validate a set of outcome and process indicators based on linked administrative health data (ie, insurance claims data linked to hospital discharge data) assessing older inpatient safety related to anticoagulation in both Switzerland and France, and enabling comparisons across time and among hospitals, health territories, and countries.

The project will thus comprise complementary steps aiming to (1) develop and validate a set of geriatric patient safety indicators (GPSIs) assessing in-hospital and postdischarge anticoagulant-related adverse events in older inpatients; (2) adapt the PSI-09 and PSI-12 to Swiss and French linked data; (3) develop a set of geriatric quality indicators (GQIs) assessing the management of anticoagulants for the prevention and treatment of arterial or venous thromboembolism based on Swiss and French data; (4) develop or adapt chronic multimorbidity indexes, proxy measures of frailty and disability, and thromboembolic and bleeding risk scores based on Swiss and French data to adjust GPSIs on case mix; and (5) compare anticoagulation safety within and between Switzerland and France.

We will develop and validate GPSIs according to the methodology used by the AHRQ [46-48], which comprises 6 standardized sequential stages (Figure 1).

Statistical analyses will be performed using Stata/MP software version 14 for Windows (StataCorp LLC) or SAS/STAT software, version 9.4 of the SAS system for Windows (SAS Institute Inc).

After accounting for inclusion and exclusion criteria, we will include 166,670 Swiss and 5,902,037 French residents aged 65 years and over in our study. Multimedia Appendix 1 comprehensively describes Swiss and French data available for our study, along with data linkage methods and success rates.

For sample size calculation, we used a test result-based sampling method [72] to minimize the number of medical records to be abstracted. We estimated optimal sample sizes, as well as sampling fractions of GPSI+ and GPSI– medical records to be abstracted, for prevalence (ie, proportion of medical records in which at least 1 adverse event was recorded) ranging from 0.4% to 5%, a desired algorithm sensitivity of 75%, a 20% width for sensitivity 95% confidence interval, and unknown numbers of at-risk stays (ie, unknown GPSI denominators). We extracted prevalence values and PPVs from the literature [16,24,32-34,44,50,73-75] and calculated sample size using Stata/IC software version 13 (see Multimedia Appendix 2). Thus, assuming a prevalence of 2%, an optimal sample of 3164 medical records should be randomly selected from participating hospitals and abstracted: 126 GPSI+ and 3038 GPSI– medical records. These results are consistent with the literature [73].

The research project was approved by the Cantonal Ethics Commission of Vaud (Switzerland) in August 2016 (Decision CER-VD 2016-00508) and by the French data protection authority (Commission National de l’Informatique et des Libertés) in March 2016 (Decision DE-2016-036).

All administrative health data are now available to the Swiss and French research teams. The research project will run until the end of 2018.

First, to our knowledge, this is the first study aiming at developing and validating a set of geriatric quality and safety indicators based on linked administrative health data, both in Switzerland and in France. Assessing the quality and safety of anticoagulation in older inpatients based on such data is also innovative. Indeed, in addition to using extended PSI-09 and PSI-12, which capture perioperative adverse events up to 30 days after discharge [22], we will develop and validate a set of GPSIs assessing thromboembolic or bleeding adverse events for various surgical procedures or medical conditions requiring curative or prophylactic anticoagulation. Similarly, our set of validated GQIs will complement the ACOVE-3 quality indicators based on administrative health data assessing warfarin prescription and surveillance in older patients with heart failure or atrial fibrillation [23]. In particular, they should provide useful information on direct oral anticoagulant management.

Second, validated GPSIs will account for an older inpatient case mix, which is crucial when comparing hospitals or health territories. Specific indexes based on Swiss and French administrative health data, including multimorbidity, polypharmacy, frailty, and disability indexes, will thus be purposefully developed or adapted from existing ones to target older inpatients’ conditions at admission [38-40,59-62]. These important predictors of adverse health outcomes (ie, mortality, morbidity, dependency, and institutionalization), health-related quality of life, and resource use [38,40,59] may be used not only as risk-adjustment factors for other health care quality and safety indicators, but also as screening tools for older patients’ vulnerability in various health care settings. Indeed, many countries, including France, try to implement frailty or vulnerability screening and management programs (eg, the French initiative “parcours de santé pour les personnes âgées en risque de perte d’autonomie” [PAERPA]) to avoid institutionalization and hospitalization, improve older citizens’ quality of life, and reduce health care costs [76,77]. Finally, assessing patient complexity and vulnerability, both at the individual and population levels, may contribute to better planning of appropriate health and social care services [76,77].

Developing or adapting individual thromboembolic or bleeding risk scores based on administrative health data is the third major strength of our study. These scores will be included in GPSI risk-adjustment models and may also contribute to assessing the quality of anticoagulant management by comparing patients’ risks with the treatment they actually received. Indeed, decisions regarding anticoagulant treatments (ie, type, dose, frequency, and duration) and surveillance should be consistent with risk-benefit assessment, especially in the older inpatient population, which is highly exposed to anticoagulant-related thromboembolic or bleeding complications. Furthermore, complementing individual scores with patient outcomes (ie, GPSIs) and data on anticoagulant management (ie, GQIs) should enrich information on effectiveness and safety of anticoagulants in “real life,” in particular direct oral anticoagulants, and eventually help improve professional guidelines [25,26,30,31].

Fourth, our project will enable comparisons of GPSIs and GQIs between Switzerland and France, which is a source of cross-country learning and performance improvement [17,18]. Indeed, despite having different organization, governance, and financing, and serving different populations, Swiss and French health systems “have similar goals and face similar challenges, such as demographic change, limited resources and rising costs” [17,18]. Between-country comparisons should thus help study differences in (1) linked administrative health data features, quality, regulation, and coding rules; (2) older inpatients’ and health care providers’ characteristics; (3) quality standards on and safety of anticoagulant management in older inpatients; and (4) health care policy and reforms toward transparency, accountability, and high-quality safe care to vulnerable older inpatients [12,15].

Our study has several limitations that should be considered when interpreting the results. First, our study populations may not be representative of all Swiss and French older inpatients. Indeed, Swiss data cover approximately 18% of the Swiss population aged 65 years and over [52], and policy holders residing in German-speaking cantons are overrepresented compared with others. French data include a very large population of nearly 6 million persons, namely 69% of the French insured aged 65 years and over. However, only salaried or retired employees are represented, while other professional subgroups and enrollees in 1 of the 16 specific insurance schemes (eg, soldiers, miners, ministers of religion, and employees of the French National Railway Company) are excluded from the study population [55]. As insured or cultural groups may differ in their risk factors, compliance with anticoagulant treatments, access to and utilization of health services, and geographic distribution (eg, employees may be underrepresented in rural areas), the internal and external validity of our study might be affected by selection bias [78]. Albeit figuring that this bias should not significantly affect our indicators, we will thus generalize our findings to the study populations and gather information on older inpatients excluded from our study. In particular, we will make a request to CNAMTS to access their data on French residents enrolled in other schemes than the general one. French data will thus cover the entire population.

Second, insurance claims data may also be incomplete or “selected.” For example, they will not include coded diagnoses on adverse events occurring in outpatient settings and will not provide information on patients’ adherence to treatment and surveillance. However, we expect to capture almost all postdischarge adverse events by screening readmissions, prescriptions or modifications of anticoagulant or antihemorrhagic treatments, outpatient procedures (eg, lower limb venous ultrasonography), and laboratory tests (eg, hemoglobin and international normalized ratio tests). Moreover, we believe that the quality of anticoagulant management is accurately assessed using large administrative databases linked over time. Swiss insurance claims may also be missing differentially according to health insurance deductibles (ie, SwF 300 to 2500 per year) [78-80]. Indeed, policyholders with the highest deductibles tend not to claim reimbursement of medical expenses when their annual amount does not reach the deductible, which leads to missing claims. However, this potential selection bias should not be significant, since health insurers estimate that only 2% to 3% of all invoices are not sent for reimbursement [52]. Finally, French data from SNIIR-AM may also lack some information regarding fully reimbursed long-term conditions. Indeed, long-term conditions are often underreported in patients fully covered for several long-term conditions or benefiting from a complementary insurance, or in nursing home residents [81]. We will overcome this limitation by deducing missing long-term conditions from medications coded in insurance claims or from diagnoses coded in hospital discharge records.

Third, hospital discharge data may also provide incomplete information. For example, data on inpatient medications will be missing for both countries, as they are not mandatory for reimbursement. We will then infer the prescribed anticoagulant treatment during hospital stay from that prescribed just after discharge using insurance claims data. Also, in Switzerland, we will not be able to identify reasons for readmission or retransfer to the same hospital within 18 days after discharge if the second stay is grouped in the same Major Diagnostic Category as the initial one. Indeed, according to the SwissDRG billing rules, such a readmission or retransfer is merged with the initial admission, leading to a single stay and discharge record. Furthermore, Swiss hospital discharge data are limited to acute care hospitals and do not cover hospital-at-home facilities, and rehabilitation or psychiatric hospitals. Thromboembolic and hemorrhagic adverse events occurring in these settings will thus be missed except severe ones that would necessitate transfer to acute care facilities.

Fourth, in both countries, neither hospital discharge data nor insurance claims data contain individual information on causes of death—which could be related to thromboembolic and hemorrhagic adverse events—or on significant factors that should be accounted for in GPSI case-mix adjustment. These factors include, for example, genetic factors, demographic characteristics (eg, ethnicity), clinical factors (eg, vital signs, results from clinical examinations or laboratory tests), health-related behaviors (eg, excessive alcohol consumption, diet, and physical activity), health literacy, and patient preferences or cultural beliefs [36].

Fifth, hospital discharge data may have potential limitations regarding coding accuracy. In particular, the quality of diagnostic and procedure coding may be affected by the quality of patient record documentation, coders’ background, training, and experience (eg, clinicians vs professional coders), coding quality controls, and unintentional and intentional coding errors (ie, “gaming” or “upcoding” to increase reimbursement) [36,37,82,83]. Similarly, coding rules and classifications, coders’ characteristics, DRG classification systems, and coding quality assurance policies (eg, coding quality controls, incentives for coding, and penalties for inappropriate coding) may vary significantly among health systems [21]. For example, Swiss and French health systems differ on coding rules for the “principal diagnosis” (ie, “condition responsible for resource use” vs “reason for admission”) [84], mean numbers of secondary diagnoses coded [21], medical coding classifications for diagnoses (International Classification of Diseases, 10th Revision [ICD-10], German Modification vs ICD-10 France) and procedures (Swiss operation classification [CHOP] vs French shared classification of medical procedures [CCAM]), DRG classification systems (SwissDRG vs French Groupes Homogènes de Malades), and coders’ profiles (professional coders vs mixed profiles, including professional coders and physicians). Limitations in coding quality may introduce systematic bias in GPSI estimates and in comparisons among hospitals, health territories, or countries, which cannot be accounted for by risk adjustment [21,36,37,84,85]. Criterion validity assessment will thus be essential to select valid GPSIs. Moreover, regarding the possible influence of coding practice to increase reimbursement, serious controls and financial penalties are in place in Switzerland to limit this issue. Similarly, serious controls are carried out in France, both at the hospital and at the Technical Agency for Information on Hospital Care (Agence Technique de l’Information sur l’Hospitalisation) levels.

Sixth, we have decided to compare the quality of anticoagulant prescription and surveillance across hospitals without adjusting GQIs for differences in case mix. However, process indicators may require case-mix adjustment when inpatients eligible to receive the related process are not perfectly specified [51] or when the “opportunity for violation of the standards” varies by case mix [37]. Indeed, guidelines related to preventive and curative anticoagulation do not account for older inpatients’ heterogeneity regarding their functional, cognitive, or social disability, health conditions, and complexity. We will thus verify that these conditions have no impact on GQI variations across hospitals and health territories.

Seventh, criterion validity assessment based on retrospective review of medical records may be less methodologically robust than assessing GPSIs based on prospectively collected data [86]. Indeed, many adverse events are not recognized during the process of clinical care without dedicated assessment, and they are thus incompletely captured in medical records [87-90]. However, major adverse events will probably be well reported in medical records, as they will contribute to the use of additional resources in the hospital, and need to be reported for appropriate reimbursement. Moreover, Klopotowska et al showed that “[adverse drug events] with evident causality and with clinically apparent and severe consequences,” such as adverse drug events resulting in hemorrhage or raised international normalized ratio, are well recognized and documented by medical teams [89]. In addition, the implementation of more valid, but more resource-intensive, studies would require important financial support that will not be easy to obtain, neither from research agencies nor from health care services.

Eighth, prior validation studies have suggested that, despite good predictive and construct validity, AHRQ PSIs demonstrate moderate sensitivity and PPVs in detecting surgical adverse events [22]. Indeed, PSI algorithms usually favor specificity over sensitivity and PPV [91]. Consequently, unless AHRQ PSI validity is improved, more robust adverse event detection methods (eg, prospective monitoring or voluntary patient safety event reporting) should be preferred for internal quality improvement, performance assessment, public reporting, and, above all, pay for performance [83,91-93]. However, recent validation studies contradict these results, at least for PSI-12 and PSI-09 [33,83,91]. Indeed, Mull et al showed that the sensitivity of a PSI-12 algorithm based on ICD, 9th Revision, Clinical Modification (ICD-9-CM) diagnosis and procedure codes was 65% (95% CI 63%-67%). He also suggested that PSI-12 sensitivity could be improved by using ICD-10-CM codes [91]. Similarly, Utter et al reported that the sensitivity of a PSI-09 algorithm based on ICD-9-CM codes could reach 85% (95% CI 67%-94%) [33]. Regarding PPVs, Winters et al found, based on a literature review and meta-analysis, pooled PPVs of 63.5% (95% CI 44.3%-82.7%) for PSI-12 and 78.6% (95% CI 73.2%-84.1%) for PSI-09 [83]. In our study, the detection of anticoagulant-related adverse events should also be improved by using hospital discharge data linked to insurance claims data. Indeed, by adapting PSI-12 and PSI-09 algorithms to linked data, Mull et al were able to capture 72% (with PSI-12) and 77% (with PSI-09) additional events occurring up to 30 days after discharge [22]. Finally, in Switzerland, this research project will be integrated into a larger research program, which will aim to develop measures of anticoagulant-related thromboembolic and hemorrhagic adverse events based on structured (ie, administrative and clinical data) and textual data. These measures should, in the end, have better sensitivity and specificity than AHRQ PSIs. The larger research program has just been funded by the Swiss National Fund.

This innovative study, which is part of a larger research program aiming to develop and validate GPSIs and GQIs in both hospital and ambulatory care settings, will provide valid and reliable outcome and process indicators to assess older inpatient safety related to anticoagulants. It should also provide new information on real-life prevention and treatment of thromboembolism, direct oral anticoagulant prescription and monitoring, and hemorrhagic adverse events related to direct oral anticoagulants in older patients. It should additionally contribute to describing older inpatients’ characteristics and health professionals’ practices in Swiss and French hospital and ambulatory care settings, and help identify geographic and temporal variations in older patient safety related to the health system. Moreover, within the frameworks of the Swiss Health 2020 policy agenda and the French National Health Strategy, the comparative assessment of hospitals and health territories using validated GPSIs and GQIs could help inform health policies, promote transparency, and improve health care efficiency.