Our data is derived from the AKUH, a leading tertiary academic hospital linked to 4 secondary hospitals based in Pakistan that treats approximately one million patients each year. Through a university-wide data initiative, coauthor ZS and team have developed processes to link and harness large datasets from the EHRs maintained in the health information management systems at AKUH.

For descriptive analyses, frequency and proportions are presented. For symptom presentation, we conducted regression analyses controlling for age (comparing those younger than 45 y and those older than 45 y) hypothesizing based on well-established evidence that clinical presentation and risk factor profile differ distinctly between these age groups. For example, women have been found to experience a lack of chest pain—a distinguishing symptom in the diagnosis of AMI—at increasing rates for those aged younger than 45 years [28,29]. We also assessed symptom differences controlling for diabetes given variations in clinical presentation and potential management inherent to those with coafflictions of disease [30]. For medication prescription, we controlled for diabetic status as well as percutaneous intervention being conducted during hospital stay as we hypothesized that these would affect clinical management and prescription medications at discharge [31]. We also controlled for admission year to account for potential changes in the burden of disease, facility capacity, or clinical practice guidelines that may have occurred during the extent of the data collection period and that might affect both the likelihood of patient presentation and clinical practice or management. We then evaluated gender differences in symptoms and medication use from model outputs using chi-square or Mann-Whitney tests as appropriate. Both unadjusted and adjusted odds ratios (OR) are presented with associated 95% CIs.

The funders of this study had no role in study design, data collection, data analysis, data interpretation, or writing of the report.

This research was determined to be exempt by the Yale Institutional Review Board as “not human subjects research” (2000026983). It was not appropriate or possible to involve patients or the public in the design, conduct, reporting, or dissemination plans of our research. This exemption is per the ethical standards of the responsible committee on human experimentation (institutional and national) and with the Helsinki Declaration of 1975.

In our study, we analyzed a total of 5358 discharge summaries, comprising 1769 women and 3589 men. The mean ages of the patients included in the dataset between 2010 to 2018 (n=5358), were 67.8 years for women and 63.3 years for men.

Combining structured and unstructured data, as described in the Methods section, AMI symptoms were detected in 3339 discharge summaries, and they were not detected in 2019 charts as described in methods (eg, due to no IHD symptoms being documented, interhospital transfer, or being critically ill on arrival). Among the 3339 patient discharge summaries with symptoms described, chest pain and shortness of breath were the predominant presenting symptoms of AMI (Table 1). Chest pain was more common in men (56.27% of women and 68.79% of men, P<.001) as was diaphoresis (5.47% of women and 7.33% of men, P=.05), differences that persisted even after regression analysis (P<.001 and .06, respectively).

Shortness of breath was detected in 41.45% (1384/3339) of patients and found to be more common among women (49.48% of women and 37.71% of men, P<.001). Symptoms considered “atypical” for AMI—or termed “noncardiac” by 2021 AHA guidelines—such as abdominal pain (7.82% of women and 4.7% of men, P<.001), nausea or vomiting (9.9% of women and 5.44% of men, P<.001), and weakness (3.58% of women and 2.5% of men, P=.08) were also more likely to be reported among women [32]. These differences persisted even after controlling for diabetes and age (Table 2), except in the case of weakness. On subanalysis, comparing people with diabetes to people without diabetes, there were statistically higher odds of weakness in people with diabetes as compared to people without diabetes (63/95 or 3.45% of people with diabetes as compared to 32/95 or 2.11% of people without diabetes, P=.03).

“Ghabrahat,” a culturally distinct term best translated as a sense of impending doom, was detected in 4.1% (138/3339) of charts, with a trend for higher prevalence in women (3.69% of men and 5.09% of women, P=.06), although not statistically significant.

We assessed medication prescriptions among a total of 4301 records that were detected in discharge summaries by the model (Figure 1); 1057 records had no AMI medication prescriptions detected. We found a total of 9 different classes of medications (Table 3) after categorizing all individual medications into these classes a priori. The most common medication prescribed overall was aspirin (n=4070, 94.6%). This was followed by statins (n=3962, 92.1%), clopidogrel (n=3659, 85.1%), and β-blockers (n=3405, 79.2%). In total, 38 patients were prescribed a clopidogrel or aspirin combination pill. Nearly half of all patients were prescribed an antidiabetic (1795/4301, 41.7%) and around a third were prescribed an ace inhibitor (1685/4301, 39.2%).

Statistically significant gender differences in medication prescriptions were detected among all medications and classes except aspirin, aspirin or clopidogrel combinations, and ticagrelor (Table 3). Prescriptions for β-blockers, clopidogrel, statins, and ace-inhibitors were more common for men in univariate analysis. Those for antidiabetics and angiotensin receptor blockers (ARBs) were more common for women. For three of four first-line medications (clopidogrel, statin, or β-blocker), women were less likely than men to be prescribed medication, in univariate analysis. In multivariable analysis, statistically significant differences persisted for the first-line medications: statins, β-blockers, and ARBs after adjusting for diabetic status, year of admission, and percutaneous intervention. Women had nearly 30% lower odds (OR 0.71, 95% CI 0.57-0.90) of being prescribed statins, and they had 40% lower odds (OR 0.67, 95% CI 0.57-0.78) of being prescribed β-blockers.

In this study, we present the novel development of an NLP tool to detect gender differences in symptom presentation and medication classifications in a dataset of patients hospitalized with myocardial infarction at AKUH in Karachi, Pakistan. The data used in our study is a subset of the dataset restricted to the last decade when a discharge summary module was incorporated into the electronic medical record system (2008-2018). To the best of our knowledge, this is the first creation of an NLP model concerning IHD in Pakistan and appears to be the first health care application of NLP in the country. We demonstrated the effective design of such a model based on an unsupervised fuzzy string-matching technique using clinical and context-specific expertise.

There is a paucity of publications using NLP in the South Asia region [11,18,33], and even fewer that applied NLP clinically in the setting of an LMIC [11,18,20,34,35]. The first report using NLP in India was published in 2016 [11]. Literature is even more sparse for reports of IHD in LMICs [18-20]. The first annotated corpus addressing CVD risk factors using data from Chinese electronic medical records was built in 2017 [20]. The corpus was constructed with the help of clinicians and linguists. Our study adds to this study by applying NLP based approach for information extraction from medical records. Further, it extended the application to address IHD data in LMICs. Promoting the expansion and usage of NLP has the potential to advance the recognition of culturally specific nuances in a variety of disease states.

It is important to highlight gender, race, and cultural nuances in the symptom presentation of AMI to improve outcomes because delays have been associated with worse morbidity and mortality in several populations [36]. Although chest pain or discomfort is the most common symptom for both men and women with AMI, there are differences in the description of the symptoms, as well as several presenting symptoms that often lead to delays at both patient as well as provider levels. The classic description of AMI described as the “Levine sign” includes chest pain, difficulty breathing, sweatiness (diaphoresis), and possible radiation to the jaw, neck, or left arm [37]. Interestingly, we found the predominant symptom for AMI was also chest pain in the Pakistani cohort as has been reported in Western cohorts, although a third of the patients did not present with chest pain [32]. These numbers are similar to those reported in a large registry of patients with AMI in the United States which reported a third of patients with AMI present without chest pain and were associated with worse outcomes [29]. Instead, we found a high prevalence of “atypical” presentations including shortness of breath, weakness, nausea, and vomiting—symptoms not classically thought to be associated with AMI nor are they taught to lay people seeking care for these complaints. Similar to the US data, we found nonchest pain symptoms to be more prevalent among women with AMI as compared to men in the Pakistani population. Our study supports focusing on discrete symptoms and moving away from the description of AMI presentation from “typical” and “atypical” symptoms to reduce gender and racial disparities. Our message is consistent with the recent 2021 society guidelines for chest pain evaluation [32].

Our gender-specific data adds to the limited literature describing characteristics of AMI in the South Asian population, a cohort at particularly high risk for adverse cardiac outcomes and where most reports are limited to types of MI, comorbidities, and procedural details without providing information on the presenting symptoms [38-40]. The few that do list symptoms do not report them by patient gender. The gap is likely due to the exhaustive work of medical record reviews that is required to abstract symptoms, as well as an overall lack of gender-based research. Our model provides an example of how NLP can fill gaps in unharmonized free text information available through medical record reviews. Not only does this add to the literature by the variants (or commonalities) we found, but also provides a unique tool for use in similar large datasets to detect gender differences, as well as other underinvestigated populations.

Our study further adds to the existing literature on LMICs by describing the presence of variations in context-specific terms such as “ghabrahat.” Specifically, the term “ghabrahat” is an Urdu term to imply “impending doom” and is otherwise quite nonspecific for AMI. It was documented in 4.1% of the sample population, with a trend toward being reported more by women. This finding is important for both public health messaging as well as to guide clinicians on the importance of such unconventional but contextually specific terms in patients with AMI in the Pakistani setting. It also demonstrates the importance of research innovation in LMIC settings where clinical outcomes will affect the specific settings where studies are being conducted.

We used NLP to also describe gender-specific differences in AMI management in Pakistan, as indicated by medication prescription. The most common medications prescribed were aspirin, statins, and clopidogrel but were uniformly underused compared to guideline-based recommendations for AMI care. This underuse was even more pronounced in women when compared to men. Only ARBs were more commonly prescribed in women, which we posit may be a reflection of a higher prevalence of heart failure. Both statins and β-blockers were more commonly prescribed for men. Similar disparities in treatment based on gender have been noted in previous studies [41,42]. Possible explanations for these differences are gender differences with women presenting less frequently with classic chest pain, higher burden of comorbidities in women, and later age at presentation for women presenting with AMI as compared to men. Sex differences in the pathophysiology of AMI, with a higher proportion of myocardial infarction without obstructive coronary arteries, could also contribute to this disparity, although not directly measured in our data [43]. Future iterative improvement can be carried out by our NLP tool to detect comorbidities and complications and for the detection of patients being undertreated. Considering that these classes of medications are also used in other CVDs such as stroke, our model can be generalized to assess medication prescription patterns in other diseases.

Our study has several limitations. The development of the model may have been subject to observation or confirmation bias based on individual biases among members of the research team. However, we feel this is unlikely to have affected the quality and effectiveness of the model, as we had purposeful diverse representation in the team. As mentioned previously, this included experts in cardiology, emergency medicine, global health, informatics, as well as representation from the consultant, fellow, resident, and medical student level, and equitable representation by gender. Moreover, there was adequate representation in our team from the study site (AKUH, Pakistan) to provide overall guidance, cross-cultural translation of concepts, and context-specific nuances during the development and validation of the tool. We feel that the rigor and diversity of the team contributed to the overall high accuracy of the model. In the future, we plan to test the model among patients with IHD in similar settings.

Naturally, due to missing data such as from incomplete charts or deceased patients discussed in methods (Figure 1), our findings are at risk for selection bias. However, the output generated by the extraction models was manually reviewed on randomly selected summaries to mitigate selection bias. Moreover, our findings are ratified by existing evidence on gender disparities in patients with IHD. Future studies may also promote the development of models that assess additional free-text clinical narratives in the medical record, drawing from triage and admission notes to provide a more comprehensive assessment of patient presentation that may be available in these sources. However, in light of these limitations, we provide compelling evidence that NLP is a valid option for harnessing data outputs from large datasets in the clinical setting from this region.

Furthermore, our study also lacks traditional expert medical record review as a reference standard for validating the proposed NLP methodology based on the fuzzy string-matching technique used to extract presenting AMI symptoms. Instead, we manually reviewed a subset of the extracted data to confirm the accuracy of the presenting symptom extraction model. This limitation highlights an area for future research, where further validation through expert review could improve the robustness of our findings.

In conclusion, the use of NLP for the identification of clinical characteristics and management of IHD is plausible in LMICs. Diverse content and regional expertise, including local clinical experts, may contribute to effective tool development for context-specific settings. Using this tool, we identified gender differences in clinical symptomatology in Pakistani patients with AMI, with women more likely to present with nonchest pain symptoms. Furthermore, we also noted gender differences with fewer first-line prescriptions being provided to women. In summary, opportunities exist for the implementation of NLP to guide clinical practice in Pakistan, and similar LMIC settings.