Microsoft word - sordoochoafinalworkshopmedinfapp032309.doc

A PSO/ACO Approach to Knowledge Discovery in a
Pharmacovigilance Context
ABSTRACT
at any dose is suspected to have resulted in adverse outcome in a We propose and evaluate the use of a PSO/ACO methodology for classification and rule discovery in the context of medication Given the limitations of premarketing trials, e.g. highly selected postmarketing surveillance or pharmacovigilance. Our study patient populations and limited duration of studies, often times considers a large data set of diabetic patients on two widely used unanticipated rare adverse events go undetected, and they only antidiabetic drugs (rosiglitazone and pioglitazone), and the risk of become more apparent when they reach the general population. myocardial infarction as an adverse effect. The goal is to Widely prescribed medications pose a substantial risk of determine the presence of previously undetected causal previously undiscovered population-level effects during relationships between therapeutics, patient characteristics, and premarketing trials. Detection of adverse events relies mainly on adverse medication outcomes. Since the proposed approach is three sources of information, namely a) data gathered from able to discover classification rules, the elicited knowledge may premarketing clinical trials; b) voluntary reporting of adverse suggest new hypotheses regarding associations between risk events in postmarketing phase and; c) information gathered from factors and an adverse event. Our classification results show high postmarketing observational studies. For example, several accuracy. Furthermore, several medication-related rules were postmarketing research studies have indicated a strong correlation discovered and analyzed. The elicited rules support previous between Cyclooxygenase-2 (COX-2) selective inhibitors, a class studies from the medical literature. Moreover, one of the studied of non-steroidal anti-inflammatory drugs (NSAIDs), with an antidiabetic drugs (rosiglitazone) was found to have a significant increase in the risk of myocardial infarction (MI) higher risk of an adverse event on diabetic, hypertensive patients, [1][5][11][16][24][29]. This was particularly true for Rofecoxib as compared to the other drug. This last finding suggests that (Vioxx) which was withdrawn from the market in September, pioglitazone may have a protective effect against myocardial infarction on diabetic, hypertensive patients. The work reported herein is part of an observational retrospective cohort study of patients on diabetic medications who may be at an Categories and Subject Descriptors
increased risk of coronary heart disease (CHD) [3]. CHD is defined as acute myocardial infarction requiring hospitalization. I.2.6 [Artificial Intelligence]: Learning – Knowledge acquisition.
We focus our analysis on two antidiabetic oral medications, I.2.8 [Artificial Intelligence]: Problem Solving, Control
Methods, and Search – Heuristic methods The purpose of the present study is two-fold: First, we seek to General Terms
evaluate potential benefits of PSO/ACO methodology as an Algorithms, Measurement, Experimentation. adjunct to more traditional statistical methods for postmarketing surveillance or pharmacovigilance (hereafter we will use both Keywords
terms indistinctively). The goal is to determine, using data from electronic medical records, the presence of previously undetected Swarm Intelligence, Ant Algorithms, PSO/ACO, Knowledge causal relationships between therapeutics, baseline characteristics Discovery, Genetic Based Machine Learning, Postmarketing (e.g. gender, race, age), comorbidities and adverse events. Surveillance, Pharmacovigilance, Healthcare. Second, use the elicited knowledge to develop potentially new hypotheses as to suspected associations between all risk factors 1. INTRODUCTION
involved that may play a critical role in the adverse outcome. The US Food and Drug Administration (FDA) defines an Adverse The remainder of this paper is organized as follows: Section 2 Drug Event (ADE) as “any incident where the use of a medication presents a brief overview of the rationale for a nation-wide and institution-wide pharmacovigilance efforts, in combination with large electronic patient databases. Section 3 describes the patient Age at time of enrollment within one of the following data used for the current study, while section 4 briefly describes 10-year intervals starting at 20 years of age: {[20-30), the PSO/ACO2 algorithm. In sections 5 and 6 we present our [30-40), [40-50), [50-60), [60-70), [70-80), [80- )}; results and findings. Finally, in section 7 we draw some conclusions and discuss possible future work. iii) Race as one of {White, Black, Hispanic, Asian, Native 2. BACKGROUND
In 2006, the Institute of Medicine (IOM) issued a report, entitled iv) Medication (Rosiglitazone / Pioglitazone); The Future of Drug Safety—Promoting and Protecting the Health v) History of cardiovascular disease (Y/N). Diagnoses of the Public [20]. Among other suggestions, the IOM report considered are coronary artery disease, angina, recommended that the FDA identify ways to access other health- congestive heart failure, cerebrovascular accident, related databases and create a public-private partnership to percutaneous coronary intervention, coronary artery support safety and efficacy studies. As a result, the FDA has been fostering public-private collaborations, leveraging increasingly available large electronic patient databases and exploring new, emerging technologies to further advance the safety and quality of vii) HBA1C as indicator for disease management (Y: all realms of healthcare. Partners Healthcare System is one of patient has been monitored/ N: patient has not been five institutions which, in conjunction with the eHealth Initiative (eHI) and the FDA, is collaborating in a nation-wide effort to viii) HBA1C > 8 as indicator of poor glycemic control and develop novel health information technology tools to create an active drug safety surveillance system across the U.S. ix) Creatinine as indicator of disease management (Y/N); Independent from nation-wide efforts, Partners Healthcare System has been carrying out patient-population pharmacovigilance x) Creatinine > 2 as indicator of chronic renal research using patient data from the Research Patient Data insufficiency and disease severity (Y/N); Registry (RPDR). Partners Healthcare System is a non-profit, xi) Hypertension (Y/N) indicated by use of any integrated health system that includes Brigham and Women's Hospital and Massachusetts General Hospital. The RPDR is a xii) Hyperlipidemia (Y/N) indicated by use of any anti- centralized data warehouse containing clinical data such as patient hyperlipidemic medications (See Table 1); demographic information, dates, medication, diagnosis information, and discharge summaries. xiii) Hospitalizations/ED visits as proxy for severity of Table 1. Characteristics table for patients on rosiglitazone or
Institutional Review Board approval was obtained prior to pioglitazone monotherapy. Values are number (%) unless
selecting a group of 2,185 diabetic patients on rosiglitazone or otherwise indicated.
pioglitazone monotherapy from a cohort study of 34,252 patients on diabetic medications. These two medications belong to the thiazolidinediones (TZs) class of oral hypoglycemic medications. Both rosiglitazone and pioglitazone were introduced to the market in 1999. Both medications have shown to increase the risk of congestive heart failure (CHF) [30], and rosiglitazone has been associated with increased risk of myocardial infarction when Only patients over 18 years of age with at least one record of prescription as an outpatient, or dispensation as an inpatient, of either rosiglitazone (n = 1594) or pioglitazone (n= 591) between January 1st, 2000 and December 31st, 2006 were included in the study. Selected patients should be on monotherapy (single antidiabetic medication) for the whole duration of the study. Our definition of monotherapy was more stringent than the definition used in the larger cohort study in order to limit confounding factors between medication intake and a possible adverse event. The outcome of the study was the incidence of acute MI (identified by ICD9 code of 410.x) requiring hospitalization. Characteristics of rosiglitazone and pioglitazone users were similar in demographics and risk factors (Table 1). Data for each patient consists of the following thirteen potential risk factors. All data are nominal values: attributes in the TS, then the created rule is not yet complete and the algorithm performs a second step where it checks those attributes with continuous values. It is worth remembering that for nominal attributes the comparison operator used is ”=” whereas, for continuous attributes both “<=” and “>” are used to define the upper and lower bounds of the range of possible values for this attribute. The best discovered rule (BestRule) is pruned and added to the rule set. An example that satisfies all the triplets <attribute 4. PSO/ACO2
operator value> in the antecedent of the rule and belongs to the In this section we present a brief overview of the particle swarm class assigned by the rule is considered correctly classified. optimization/ant colony optimization (PSO/ACO2) algorithm These correctly classified examples are removed from the TS. proposed by Holden and Freitas [17]. Both PSO and ACO The iteration process continues until the number of unclassified algorithms mimic a population of decentralized, self-organized examples for the current class C falls below a predefined individuals that collectively work towards finding best solution(s) threshold (MaxUncovExampPerClass). Once this threshold is through an iterative searching process. Convergence to an reached, all the removed training examples are returned to TS, and optimal or near optimal solution is reached by social interaction the algorithm continues execution for the next class Ci. amongst individuals, either by exchanging information with local This section presented a brief description of the PSO/ACO neighbors –in the case of particles in PSO- or by updating a algorithm. For a more comprehensive description, see [17][18]. pheromone trail –in the case of ants in ACO. PSO/ACO2 has been mainly used to discover classification rules 5. RESULTS
in the context of data mining. This algorithm is capable of For our experiments, we used a freely available Java handling nominal attribute values without converting them into implementation of the PSO/ACO2 v1.0 rule induction algorithm numbers, as well as continuous data values. [19] and the previously described data set of 2,185 diabetic patients on rosiglitazone or pioglitazone monotherapy (section 3). We used the standard 10-fold cross validation, precision fitness function, PSO continuous optimizer, and 200 iterations for all four experiments. The only varying parameter between add all training examples to TS experiments was the number of particles, which was set to 102, WHILE (number of uncovered examples belonging to 152, 202 and 252 for each experiment respectively. Figure 1. Sequential algorithm used by the PSO/ACO2 for
It can be seen from Table 2 that the classification accuracy for the knowledge discovery – in pseudocode (from [17][18]).
PSO/ACO2 algorithm is very similar for all four configurations, with 20^2 being slightly better than the rest in terms of avg. class C > MaxUncovExampPerClass) classification rate and standard deviation. Discover best nominal rule Rule for the class C When applicable, add continuous terms to Rule Table 2. Classification accuracy on diabetes dataset for the
Return best discovered rule BestRule PSO/ACO2 algorithm with varying number of particles.
Particles
10^2 15^2 20^2 25^2
BestRule to rule set: RS = RS ⋃ BestRule Update TS by removing correctly classified Accuracy
(Avg±SD)
TS = TS – {correctly classified examples} Given the context of our study, we focused our analysis of elicited knowledge on medication-related rules, since one of our goals is Order rules in RS by descending quality to elucidate whether a) there are causal relationships between Discovered knowledge is represented in the form of rules, where therapeutics (medications), comorbidities, and baseline each rule consists of a set of one or more <attribute operator characteristics, and adverse events and b) we can detect such value> triplets (antecedent) and a consequent indicating the class signals with the PSO/ACO2 algorithm. The next section presents to which the classified object belongs. For nominal values, the operator used is “=”, whereas for continuous attribute values “<=”, “>” are used: 6. KNOWLEDGE DISCOVERY
IF <attrib op value> AND…AND <attrib op value> THEN <class> In this section we present our analysis of discovered rules using the results produced by PSO/ACO2 with 202 particles. We PSO/ACO2 algorithm depicted in Figure 1 (based on [17][18]) analyzed all medication-related rules found. These rules are listed carries out the knowledge discovery process starting with an empty rule set (RS), sequentially searching the space of possible solutions to discover one classification rule at a time as follows: We validated the discovered rules by a) providing references to For each class C, the algorithm iterates through a set of training literature supporting similar findings and/or b) performing a crude examples (TS) from which rules will be created. In a first step, relative risk analysis of variables in each rule. only rules with nominal attributes are evaluated and the The relative risk (RR) estimates the magnitude of an association discovered rule (Rule) is returned. If there are continuous between potential factors and an adverse event. It measures the incidence of the event in the exposed group compared with the proxy for severity of disease, so having no visits indicates that the non-exposed group. A relative risk of 1.0 indicates that the patient is relatively healthy for his/her condition and there are no incidence rates in both groups are identical and there is no contributing factors that may increase the risk of an adverse event. association between the potential factors and the outcome. A The fourth rule Table 3: “If Medication = P and PriorMI = N relative risk of less than 1.0 indicates a negative association, or then no_event” indicates that a patient on pioglitazone is at lesser protective effect between potential factors and the outcome under risk of having and adverse event if there is no prior myocardial study, while a relative risk greater than 1.0 indicates a positive infarction (MI). A myocardial infarction may compromise the association or an increased risk of an adverse event [15]. function of the heart and may increase the risk of subsequent In the following section we will see that some of the causal events. This is particularly true in patients with diabetes [6] [13]. relationships have a truly protective effect, that is, a relative risk Patients on pioglitazone with evidence of having an MI prior to of less than 1.0 (rules 1-3 in Table 3, analyzed in section 6.1.1), the study had a 2.68 (CI 1.61 – 4.46) risk of having an event when while others, may show a protective effect in comparison – as in compared to patients on pioglitazone who did not have an MI the case of rule 5 in Table 3, analyzed in section 6.1.2. prior to Jan 1st, 2000. For patients on rosiglitazone there is a slightly higher risk of having an MI if the patient has had a prior Table 3. List of discovered medication-related rules.
Medication = R indicates Rosiglitazone, and
Medication = P indicates Pioglitazone
6.1.2 Rosiglitazone-Related Rules Rules five to eight in Table 3 depict causal relationships between If Medication = P and hasHBA1C = Y then no_event rosiglitazone and an adverse event. Rule five “If Medication = R If Medication = P and Hospitalizations/ED = N then no_event and Hypertension = N then no_event” indicates that a patient on Rosiglitazone with no hypertension may have a lesser risk of If Medication = P and PriorMI = N then no_event having an adverse event. Our calculations indicate that If Medication = R and Hypertension = N then no_event hypertensive patients on rosiglitazone have a relative risk of 9.90 If Medication = R and PriorMI = N and hasCreatGt2 = N (CI 8.46 – 11.60) of having an adverse event, compared to hypertensive patients on Pioglitazone (RR 2.87; CI 2.02 – 4.07). If Medication = R and Gender = M and Age_Range = 50-60 Given the fact that coronary artery disease and hypertension are common risk factors in patients with diabetes [14][22], patients If Medication = R and Age_Range = 60-70 and presenting these conditions may be more susceptible of having an 6.1 Analysis of Discovered Rules
Rule six “If Medication = R and PriorMI = N and hasCreatGt2 = N then no_event” indicates that patients on rosiglitazone with no prior MI and creatinine levels within normal values are at There are four pioglitazone-related rules (Table 3). The first rule considerably lower risk of having an event (RR = 0.44; CI 0.37 – refers to the administration of pioglitazone with attributes taken 0.53) when compared to patients on rosiglitazone with abnormal into consideration. The rule suggests no association between creatinine levels (RR 2.253; CI 1.828 – 2.777)). [1][23] indicate pioglitazone and the possibility of an adverse event (MI). Our that abnormal kidney function increases the risk of myocardial calculations indicate that the relative risk (RR) of having an adverse event if a patient is on pioglitazone compared to a patient Rule seven “If Medication = R and Gender = M and Age_Range on rosiglitazone is 0.6086 (Confidence Interval (CI) 0.467 – = 50-60 then has_event” indicates that a 50-60 y/o male patient 0.7933). This is consistent with reports from [8] indicating that on rosiglitazone is slightly more likely to have an adverse event pioglitazone may have a neutral to favorable effect towards (RR = 1.735; CI 1.155 – 2.606) than a female patient in the same cardiovascular adverse events. Similarly, [10] and [31] have age group. This is consistent with reports from [2] indicating that reported that rosiglitazone may have a higher risk of although both diabetic men and women are at higher risk of an cardiovascular events compared to pioglitazone. adverse cardiovascular event than non-diabetic patients, diabetic The second rule in Table 3: “If Medication = P and hasHBA1C = men in this age range are at a higher risk than diabetic women. Y then no_event” indicates that if a patient is taking pioglitazone The last rule in Table 3 “If Medication = R and Age_Range = (Medication = P) and patient’s HBA1C has been monitored 60-70 and hasHBA1C = Y then has_event”. The rule itself (hasHBA1C = Y) then there is no event. It is worth remembering indicates that patients within this age range and with HBA1C that HBA1C is used as a proxy for glycemic control and disease monitored may have an event. Our calculations indicate that management. hasHBA1C = Y indicates that the patient’s blood patients on rosiglitazone within this age range and with monitored sugar has been monitored. A relative risk of 0.65 (CI 0.42 – 0.99) HBA1C do not seem to be particularly at risk of having an event indicates that a patient on pioglitazone with monitoring of (RR= 0.70; CI 0.517 – 0.948) when compared to all patients on HBA1C is less likely to have an adverse event than patients on rosiglitazone and HBA1C = Y (RR= 0.769; CI 0.639 – 0.925). pioglitazone with no monitoring of HBA1C [21] [26]. This suggests that: a) patients in this particular age range may not The third rule in Table 3: “If Medication = P and be at a higher risk of an adverse event and; b) appropriate Hospitalizations/ED visits = N then no_event” indicates that a glycemic monitoring may reduce the overall risk of an adverse patient on pioglitazone with no hospitalizations or emergency event and improve disease management [25]. department (ED) visits is less likely to have an event with a RR = 0.423; CI 0.269 – 0.665). Hospitalizations/ED visits is used as a 6.2 Visualizing Causal Relationships
7. DISCUSSION AND FUTURE WORK
We explored the applicability of a heatmap to visualize causal Overall, elicited causal relationships between therapeutics, patient relationships in an easy-to-interpret manner. Similar to a weather characteristics and an adverse event were consistent with findings map, where temperature is encoded by color, in a heatmap we in medical literature. These findings support our initial depict the potential risk of an adverse event given specific assumptions as to the suitability of PSO/ACO as an alternative to combination of factors (e.g. medication, comorbidities, baseline more traditional methods for knowledge discovery in a characteristics) in terms of ‘temperature’, where ‘cold pharmacovigilance context. Furthermore, due to the inherent temperatures’ represent a low relative risk, and ‘hot temperatures’ nature of the applied algorithm, elicited rules were seamlessly represent a high relative risk of an adverse event. coupled into a visual display easy to understand, thereby We display potential causal relationships in a two-dimensional increasing the applicability and understandability of these heatmap where the color of a cell in the x,y position depicts the relative risk for patients on medication x who had a comorbidty y We expect that the elicited knowledge may provide critical of having an adverse event. The first two rows in Figure 2 depict insight into potentially worrisome combination of factors that the relative risk of having an adverse event for patients on may increase the risk of an adverse event. For example, depicted rosiglitazone (R) and pioglitazone (P). The third row depicts the in Figure 2, column five, we see that diabetic hypertensive risk of having an event regardless of the medication (overall). For patients on rosiglitazone are at higher risk for an adverse event example, the cell in position row 2, column 3 in Figure 2 is a while pioglitazone seems to have a protective effect on diabetic color-coded representation of the (low) relative risk of having an hypertensive patients. This observation prompts two hypotheses adverse event for patients on pioglitazone who had HBA1C levels worthy of further investigation: a) is this increased risk due to the monitored, as detected by rule 2 in Table 3. fact that thiazolidinediones in general can among other things, increase fluid retention, and hence increase blood pressure [22]? Is this particularly true for rosiglitazone but not for pioglitazone? Or b) could this be due to a possible drug-drug interaction between rosiglitazone and antihypertensive drugs? In Summary, our findings are by no means exhaustive, but demonstrate that the potential benefits of PSO/ACO for knowledge elicitation are many-fold: The approach itself is capable of discovering causal relationships in the form of rules from patient data extracted from electronic medical records; discovered rules can be easily mapped into heatmaps – or any other visual aid - to provide users with ‘at-a-glance’ immediate interpretation of findings; rules could be seamlessly incorporated into monitoring systems [12]; elicited knowledge may serve to develop new hypotheses as to suspected associations between all risk factors involved that may play a critical role in the adverse Figure 2. Heatmap depicting relative risk of an adverse event
Important directions for future work include: a) extending our given a risk factor for patients on rosiglitazone (R),
analysis of discovered rules; b) further investigate findings of pioglitazone (P) and regardless of medication (overall).
diabetic hypertensive patients and their use of antihypertensive medications; c) improve visual display of results; d) explore other Although exploratory, we have found this strategy of representing knowledge discovery methods and compare results and; e) extend the strength of causal relationships by colors extremely powerful. our model to include other possible adverse events. Since the color of a cell depicts a quantitative risk relative to other cells in the same column, it is possible for users to identify 8. ACKNOWLEDGEMENTS
potential trends and outliers in data. For example, column 5 in We would like to thank Drs. Mena Macedo and Margarita Figure 2 depicts the risk of an adverse event in hypertensive Sanchez for their advice in writing this manuscript. patients on rosiglitazone (row 1), pioglitazone (row 2) and overall (row 3). It shows that hypertensive patients on rosiglitazone may have a potentially higher risk of having an adverse events 9. REFERENCES
compared to hypertensive patients on pioglitazone. Further, since [1] Beddhu S, Allen-Brady K, Cheung AK, Horne BD, Bair T, hypertensive patients on pioglitazone seem to have a lower Muhlestein JB and Anderson JL. Impact of renal failure on relative risk, this could be interpreted as pioglitazone having a the risk of myocardial infarction and death. Kidney protective effect against myocardial infarction on hypertensive International (2002) 62, 1776–1783. [2] Booth GL, Kapral MK, Fung KT, Jack V. Relation between This section presented the analysis of medication-related elicited age and cardiovascular disease in men and women with rules. In the following section we analyze our findings and diabetes compared with non-diabetic people: a population- based retrospective cohort study. The Lancet 2006; 368(9529):29-36. [3] Brownstein JS, Murphy SN, Goldfine AB, Grant RW, Sordo [15] Hennekens CH, Buring J. Epidemiology in Medicine. Little M, Gainer V, Colecchi J, Dubey A, Nathan DM, Glaser JP, Kohane IS. Real-time surveillance of cardiovascular risk [16] Hippisley-Cox J, Coupland C (2005) Risk of myocardial associated with diabetic medications from electronic medical infarction in patients taking cyclo-oxygenase-2 inhibitors or conventional non-steroidal anti-inflammatory drugs: [4] Brownstein JS, Sordo M, Kohane IS, Mandl KD. The tell- population based nested case-control analysis. BMJ 330: tale heart: population-based surveillance reveals an association of rofecoxib and celecoxib with myocardial [17] Holden N, Freitas AA. A Hybrid PSO/ACO algorithm for infarction. PLoS ONE 2007;2(9):e840. classification. Proceedings of the 9th Genetic and [5] Caldwell B, Aldington S, Weatherall M, Shirtcliffe P, Evolutionary Computation Conference Workshop on Particle Beasley R (2006) Risk of cardiovascular events and Swarms: The Second Decade (GECCO ’07), pp. 2745-2750, celecoxib: a systematic review and meta-analysis. J R Soc [18] Holden N, Freitas AA. A Hybrid PSO/ACO Algorithm for [6] Canto JG, Shlipak MG, Rogers WJ, Malmgren JA, Frederick Discovering Classification Rules in Data Mining. Journal of PD, Lambrew CT, Ornato JP, Barron HV, Kiefe CI. (2000). Artificial evolution and Applications. Vol 2008, Article id Prevalence, clinical characteristics, and mortality among patients with myocardial infarction presenting without chest [19] Holden N, Freitas AA. PSO/ACO2 Java application. pain. JAMA. 2000;283(24): 3223–9. Downloads at http://sourceforge.net/projects/psoaco2/ [7] Carvalho DR, Freitas AA, Ebecken N. Evaluating the [20] Institute of Medicine, The Future of Drug Safety- Promoting Correlation Between Objective Rule Interestingness and Protecting the Health of the Public, September 22, 2006, Measures and Real Human Interest. PKDD 2005, LNAI [21] S. B. King III. Acute myocardial infarction: are diabetics [8] Dormandy JA, Charbonnel B, Eckland DJ, et al. Secondary different? J. Am. Coll. Cardiol., May 1, 2000; 35(6): 1513 - prevention of macrovascular events in patients with type 2 diabetes in the PROactive Study (PROspective pioglitAzone Clinical Trial In macroVascular Events): a randomised [22] Nesto R, Libby P. Diabetes mellitus and the cardiovascular controlled trial. Lancet 2005;366(9493):1279-89. system. In: Braunwald E, Zipes DP, Libby P, eds. Heart Disease: A Textbook of Cardiovascular Medicine. 6th ed. [9] FDA issues public health advisory on Vioxx as its Philadelphia, Pa: W.B. Saunders Co; 2001: 2133–2150. manufacturer voluntarily withdraws the product. FDA News. 2004 Sep 30. [23] Nichols GA, Hillier TA, Erbey JR, et al. Congestive heart http://www.fda.gov/bbs/topics/news/2004/new01122.html failure in type 2 diabetes: prevalence, incidence, and risk factors. Diabetes Care. 2001; 24: 1614–1619 [10] Gerrits CM, Bhattacharya M, Manthena S, Baran R, Perez A, Kupfer S. A comparison of pioglitazone and rosiglitazone for [24] Nissen SE, Wolski K. Effect of rosiglitazone on the risk of hospitalization for acute myocardial infarction in type 2 myocardial infarction and death from cardiovascular causes diabetes. Pharmacoepidemiol Drug Saf 2007;16(10):1065- [published correction appears N Engl J Med. 2007;357(1): 100]. N Engl J Med 2007;356(24):2457-2471. [11] Graham DJ, Campen D, Hui R, Spence M, Cheetham C, et [25] Qaseem A, Vijan S, Snow V, Cross T, Weiss KB, Owens al. (2005) Risk of acute myocardial infarction and sudden DK. Glycemic Control and Type 2 Diabetes Mellitus: The cardiac death in patients treated with cyclo-oxygenase 2 Optimal Hemoglobin A1C Targets. A Guidance Statement selective and non-selective non-steroidal anti-inflammatory from the American College of Physicians. Ann Inter Med drugs: nested case-control study. Lancet 365: 475–481. [12] Greenes RA, Sordo M, Zaccagnini D, Meyer M, Kuperman [26] Singer DE, Nathan DM, Anderson KM, Wilson PW, Evans GJ. Design of a Standards-Based External Rules Engine for JC. Association of HbA1c with prevalent cardiovascular Decision Support in a Variety of Application Contexts: disease in the original cohort of the Framingham Heart Report of a Feasibility Study at Partners HealthCare System. Study. Diabetes, Vol 41, Issue 2 202-208. Proceedings of Medical Informatics Association (MedInfo), [27] Singh S, Loke YK, Furberg CD. Long-term risk of cardiovascular events with rosiglitazone: a meta-analysis. [13] Haffner SM, Lehto S, Ronnemaa T, Pyorala K, Laakso M. Mortality from coronary heart disease in subjects with type 2 [28] Solomon DH, Schneeweiss S, Glynn RJ, Kiyota Y, Levin R, diabetes and in nondiabetic subjects with and without prior et al. (2004) Relationship between selective cyclooxygenase- myocardial infarction. N Engl J Med 1998;339(4):229-34. 2 inhibitors and acute myocardial infarction in older adults. [14] He J, Ogden LG, Bazzano LA, et al. Risk factors for congestive heart failure in U.S. men and women: NHANES I [29] Solomon SD, McMurray JJ, Pfeffer MA, Wittes J, Fowler R, epidemiologic follow-up study. Arch Intern Med. 2001; 161: et al. (2005) Cardiovascular risk associated with celecoxib in a clinical trial for colorectal adenoma prevention. N Engl J Med 352: 1071–1080. [30] Yki-Jarvinen H. Thiazolidinediones. N Engl J Med. 2004; With Diabetes Who Initiated Rosiglitazone vs Pioglitazone Therapy. Arch Intern Med 2008;168(21):2368-75. [31] Winkelmayer WC, Setoguchi S, Levin R, Solomon DH. Comparison of Cardiovascular Outcomes in Elderly Patients

Source: http://www.cs.stir.ac.uk/~goc/papers/GECCO09SordoOchoaFinalWSsubmitted032309.pdf