COVID-19 impacted the perceived quality of all facets of health care services across many European countries [1]. The pandemic led to the postponement of nonurgent appointments and treatments [2], prioritization of COVID-19 cases, and patient-initiated cancellations due to fear of infection [3]. Consequently, primary health care use and outpatient visits declined significantly, with a reported 60% decrease in hospital outpatient visits in the United States by early April 2020 [4].

Although the pandemic was officially declared over on May 5, 2023 [5], its effects continue to impact health care services. Physical distancing measures influenced service delivery, distinguishing between procedures requiring physical intervention (eg, surgery) and consultative services feasible through telemedicine (eg, psychotherapy) [5]. The backlog of postponed treatments and the rapid adoption of telemedicine have hindered full system recovery [1,5].

All aspects of health care services have been influenced by the pandemic. However, in this research, we specifically focused on primary health care services. Primary health care centers serve as the first point of contact for patients and play a main role in community health. Given their important role, it is essential to examine how the pandemic has impacted the perceived quality of care and whether these changes have persisted over time.

While it is important to study the internal functioning of the primary health care system, it is also important to study the quality of the primary health care system from the point of view of the patients. It is known that patients’ satisfaction regarding primary health care service quality is an important indicator of the success of the treatment [1] and of the overall quality of health care [2]. Patient satisfaction is typically monitored via self-report questionnaires [3]. Sometimes, qualitative methods such as interviews are also used. These are pertinent methods, but to complement them, we propose a methodology in which freely available user-generated content (UGC) can be exploited to gain additional insights into patient satisfaction. User-generated data from Google Maps in the form of user reviews can be mined to assess user satisfaction regarding health care services over time.

Research using Google Maps review data for analyzing user sentiment regarding primary health care is scant. The work by Lee et al [4] presents an analysis of Google Maps reviews of one hospital in Taiwan. As stated in the work by Ruelens [5], the amount of UGC on this topic is increasing, in part due to the COVID-19 pandemic, but is yet to be explored to its full potential. This study aimed to use this resource and further investigate its potential. While there is research on exploiting UGC from other platforms such as Facebook, X (formerly known as Twitter), and discussion forums, research using Google Maps reviews to monitor user sentiment in health care is scant. However, exploiting all types of UGC in addition to user surveys is valuable because user-generated reviews may highlight the views of those users who typically do not answer surveys. Other benefits of exploiting UGC is that, once the sentiment monitoring system is built and running, it can be operated cost-effectively to monitor user sentiment in real time. Sentiment analysis can also address issues that are not asked in a survey but that are expressed in the free-form writing of the reviews, thus providing a potentially richer insight into customer sentiment. Moreover, analysis of textual data collected through surveys often involves expensive labor-intensive manual coding, and the results are typically reported at longer intervals. In addition, UGC may reflect the views of users while they are being served and not after the service, which is the case with surveys [4]. To the best of our knowledge, this is the first study to use sentiment analysis to compare opinions obtained from Google Maps reviews before and after the COVID-19 pandemic. Our study is also the first attempt to assess user sentiment in Finland and Spain before, during, and after the COVID-19 pandemic; therefore, this is also the first study that compared the opinions in 2 different countries with different regulations on the COVID-19 pandemic. To bridge the gap in patient sentiment analysis research addressing perceived service quality in primary health care before, during and after COVID-19 using user generated data, we addressed the following research questions:

This research aimed to investigate the changes in public sentiment regarding primary health care in Finland and Andalusia (Spain) before, during, and after the COVID-19 pandemic using UGC from Google Maps reviews. The second goal of this study was a methodological one as we also attempted to assess how to build such a data processing pipeline. The results are expected to contribute to the body of knowledge on both the impacts of the COVID-19 pandemic on health care services and the methods for collecting and analyzing data regarding perceived service quality in primary health care.

The number of COVID-19 deaths was chosen as a comparative metric between both countries as it provides a relatively clear outcome. It offers insight into the pandemic’s severity, but it should be interpreted with caution as we need to acknowledge that different countries adopted different approaches on the recording and reporting of deaths attributed to COVID-19. Moreover, these practices evolved over the course of the pandemic.

Figure 1 shows the number of deaths per million over time for the 2 locations: Andalusia (in orange) and Finland (in blue). The x-axis represents the dates, whereas the y-axis represents deaths per million. Each data point represents approximately a 20-week period to offer insights with enough granularity into how the situation progressed over time in both locations. In the case of Finland, as COVID-19 data were available on the National Institute for Health and Welfare [24], the number of deaths was accounted for until the day of the last collection of Google Maps reviews (in May 2024). On the other hand, for Andalusia, data were included up to the point when they were no longer collected by the authorities (June 2023). In both cases, the start date of the data collection was January 1, 2020. To account for population size differences, the deaths per million people were calculated by dividing the absolute death numbers by the population of each location (8.5 million for Andalusia and 5.5 million for Finland).

In the case of Andalusia, the data regarding the number of deaths were collected from the National Center for Epidemiology [25,26]. Although the registration of new data stopped in July 2023, the existing data allowed for tracking the evolution of the pandemic in Andalusia.

In the case of Finland, the National Institute for Health and Welfare’s Finnish National Infectious Diseases Register reports ongoing key information on COVID-19.

As can be observed in Figure 1, both countries experienced multiple death peaks but with different temporal distributions. The number of deaths per million in Andalusia over the time frame reveals that this region experienced more severe and frequent spikes in mortality rates, especially in 2021. Finland shows lower peaks.

To understand the similarities and differences in perceived service quality of primary health care, country-specific COVID-19 policies and their timing in Finland and Andalusia should also be considered. There are cross-country frameworks and datasets for measuring government response to COVID-19, such as the Oxford COVID-19 Government Response Tracker [27] and COVID-19 Health Systems Response Monitor [28,29]. They have been used in various cross-country analyses, such as comparing responses to COVID-19 [30]. There are also country- or region-specific frameworks such as the severity index for assessing nonpharmaceutical interventions on COVID-19 transmission in Spain [31]. These frameworks and indexes aggregate various response measures, including restrictions on public gatherings, school closures, curfews, and the maintenance of essential health services.

To compare the policies in Finland and Andalusia, the COVID-19 stringency index suggested in the Oxford COVID-19 Government Response Tracker [32] may be used. It is a composite measure of 9 response metrics: school closures, workplace closures, cancellation of public events, restrictions on public gatherings, closures of public transport, stay-at-home requirements, public information campaigns, restrictions on internal movements, and international travel controls. The data for the stringency index were collected at the national level for Finland and Spain [33] from January 1, 2020, to December 31, 2022, and they are shown in Figure 2. The figure shows that the response measures were stricter in Spain almost throughout the COVID-19 pandemic, with an exception at the end starting from November 2022.

The resulting table for review data collection had the following structure: place_id, place_name, place_location_link, review_datetime_utc, review_timestamp, review_rating (number of stars on Google Maps), review_likes, review_text, review_link, and region. Table 2 shows an example of a review for one of the medical centers in Finland. This review was posted on September 15, 2023, at 6:48 PM. The user’s rating of the review was 5. Most of the review texts from Finland were in Finnish, whereas for Andalusia, they were in Spanish; however, part of the reviews were in other languages. The organizations with the highest number of reviews with text in Finland were Helsinki University Hospital Myllypuro Laboratory (n=153 reviews), Kontinkangas Health and Well-being Centre (n=228), and Kalasatama Health and Well-being Centre (n=199), whereas in Andalusia, they were the Limonar Health Center (n=328), Dr Joaquín Pece Health Center (n=281), and Las Lagunas Health Center (n=283).

The data extracted from Google Maps using Outscraper were stored in a tabular format. First, we filtered out the non–Spanish-language text reviews for Andalusia and the non–Finnish-language text reviews for Finland using langdetect [24]. A total of 7.9% (2195/27,756) non–Spanish-language reviews and 13.9% (913/6558) non–Finnish-language reviews were removed, resulting in 25,561 reviews for Andalusia and 5645 reviews for Finland.

Initially, to conduct the sentiment analysis, 2 tools were considered: pysentimiento [26] and AFINN [41]. pysentimiento is a Python toolkit; the model is transformer based; and it can analyze text in Spanish, English, Italian, or Portuguese. To add a new language to this toolkit, it is necessary to train a new model. AFINN [42] is a wordlist-based approach to sentiment analysis that uses preannotated dictionaries where each word is assigned a sentiment score between –5 and 5. A score of 0 means a neutral sentiment, a negative score means a negative sentiment, and a positive score means a positive sentiment. The AFINN tool produces as output the sum of the scores of the emotionally loaded words in the text. Thus, for example, a long sentence with many negatively loaded words will receive a lower score than a short sentence with proportionally as many negative words. This scoring reflects the fact that, if a user makes the effort to write a long and negative review, it is given more importance than a short review with proportionally as many negative words.

As a disclaimer for scientific reproducibility, at the time this research was being conducted, the version available on the Python Package Index repository did not include the Finnish language, but the one on the official GitHub repository did. Furthermore, AFINN offers a Spanish dictionary available on GitHub.

Finally, pysentimiento was discarded for 2 primary reasons: first, the analysis could not be conducted for the Finnish language, and second, its output format was incompatible with that of the AFINN tool.

Deep learning–based methods were applied for confirming the results obtained through AFINN. The following fine-tuned deep transfer learning models were used: bert-finnish-sentiment-analysis-v2 [43] for Finnish and twitter-XLM-roBERTa-emotion-es [44,45] for Spanish. The Finnish sentiment analyzer is a fine-tuned version of the uncased bidirectional encoder representations from transformers deep transfer learning model for Finnish [46]. The fine-tuning was conducted using the nisancoskun/finnish_sentiment_data dataset [47]. The sentiment analyzer classified the reviews into positive and negative. The Spanish sentiment analyzer was based on the multilingual XLM-roBERTa-base model [48] that has been fine-tuned for emotion classification in Spanish. It classified the reviews into the following 7 emotions: anger, disgust, fear, joy, sadness, surprise, and other, the latter representing other emotions or the absence of any emotion. In our analysis, we regarded the emotions anger, disgust, fear, and sadness as negative and joy and surprise as positive.

The entire dataset was organized into a DataFrame and subsequently divided into 3 distinct subsets: one representing the period before the COVID-19 pandemic, one representing the period during the pandemic, and one representing the period after the end of the pandemic had been declared [49]. This was the first step of our pipeline, as can be seen in step DC1 in Figure 3. Table 3 shows the date ranges of the DataFrames.

In this context, the decision to divide the data into 3 temporal phases (before, during, and after the pandemic) not only responds to an obvious chronological logic but is also widely justified from a methodological, social, and health care perspective. This segmentation allows for a more accurate capture of variations in public perceptions of the health care system, as well as changes in the nature of interactions among patients, health care professionals, and institutions.

Before the pandemic, the health care system operated under a relatively stable paradigm. In-person consultations were the norm, and user ratings tended to focus on classic aspects such as waiting times, quality of care received, and administrative efficiency. During this period, sentiment analysis reveals a more neutral or predictable emotional distribution influenced by structural or regional factors but without major external changes that distort general perceptions.

The outbreak of the pandemic caused a profound disruption in the organization and functioning of health care systems worldwide. During this phase, there was an unprecedented overload of hospital services, as well as the suspension or postponement of numerous nonurgent consultations and procedures. Furthermore, the use of alternative care channels such as telemedicine, which until then had been marginally used, became widespread. Consequently, the emotional content of opinions on and assessments of medical care during this period is qualitatively different. Fear, uncertainty, frustration over the lack of resources, and gratitude toward health care professionals emerge as dominant emotions, making it essential to analyze this period separately to avoid distortions in the analysis models.

The period after the pandemic, while not implying a full return to prepandemic conditions, marks a new scenario characterized by a re-evaluation of the lessons learned from the health care system. The population expresses new expectations regarding the quality and accessibility of services, as well as a higher level of demands on institutional management. Similarly, new topics emerge in patient reviews and opinions, such as the aftereffects of post–COVID-19 condition, the impact on mental health, and the accumulation of waiting lists. This phase also allows us to observe how the social perception of the health care system underwent a lasting transformation, reflecting a heightened awareness of its vulnerability and the need for resilience.

From a methodological perspective, temporal segmentation is essential for correctly interpreting the data. The meaning of certain expressions changes radically depending on the context. For example, a statement such as I was not seen in the emergency room can have very different implications if it is uttered before the pandemic, during the hospital collapse, or during the recovery phase. Similarly, the volume of reviews or publications may have skyrocketed during the pandemic period driven by the emotional burden and visibility of the health care system, requiring differentiated statistical treatment.

Finally, this temporal division offers added value from the perspective of health care management and public policy. It allows for clearly identifying critical moments in the system, evaluating the effectiveness of the measures adopted, and drawing lessons applicable to future crises. Furthermore, it facilitates the design of more effective communication strategies aligned with the emotions and expectations of the population at each phase.

Online reviews reflect patient experience with certain delay. However, most reviews are written within the first week of the patient experience: “Nearly half (47%) of patients are most likely to leave a review within 24 hours of their appointment, and 74% said they would be more inclined to do so if their provider asked them to” [50]. Thus, as the periods for analyzing sentiment trends in our research were long (more than a year) and most reviews are written within a week, the influence of this delay was insignificant for this analysis and was not considered when we created the subsets.

After the datasets had been selected, the UGC (Google Maps reviews) was analyzed (step DA2.1 in Figure 4). Analysis of the most frequent words and nouns associated with positive and negative review sentiment was also conducted. A simple categorization of the reviews was proposed dividing positive and negative ones: reviews above and including a rating of 4 stars were considered positive, whereas reviews below and including a rating of 2 stars were understood as negative. Language-specific text processing steps were applied to the data. For this purpose, we used natural language processing models from the Python spaCy library [51]. The pipeline involved tokenization, lemmatization, stop word removal, and filtering of punctuation. The most common words and nouns were extracted from the processed tokens. After this first step, we removed reviews without text, and we conducted the same analysis (step DA2.2). Figure 4 shows all the steps taken in the processing of the data.

Following the guidelines of the ethics committee of Haaga-Helia University of Applied Sciences, an institutional ethical review was deemed unnecessary for this study, which collected and analyzed publicly available social media data. All presented reviews were translated into English and anonymized to avoid association with any particular user of the Google Maps platform.

As described in the previous section, the first analysis captured the frequency distribution of user ratings (step DA2.1 in Figure 4). Google Maps ratings range across a scale from 1 to 5. Figure 5 presents the aggregated data from Finland across the defined phases—before the pandemic, during the pandemic, and after the pandemic—and the cumulative data from all 3 periods. From the prepandemic period, a total of 6294 reviews were analyzed, with a mean rating of 3.48 (SD 1.489) and a median rating of 4.0 (IQR 3), which indicates a generally positive user sentiment. During the pandemic, 4026 reviews were processed, with a mean rating of 3.31 (SD 1.661) and a median rating of 4.0 (IQR 4), reflecting a similar trend in user feedback. From the postpandemic period, we analyzed 1932 reviews, with a mean rating of 3.28 (SD 1.826) and a median rating of 4.0 (IQR 4; Table 4).

Figure 5 allows us to observe that the plots’ shapes have the same form. The values for the median and mean ratings in Table 4 confirm what can be visually observed in the figure—the numeric user ratings did not show noteworthy changes in Finland. In Table 4, we observe that the median stayed the same and the mean showed a very slight decrease.

Before conducting the sentiment analysis on the textual reviews, we took out the reviews without written text or not written in Finnish or Spanish (step DA2.2). The data regarding the Finnish reviews with text are shown in Figure 6. In the figure, we included the information from before the pandemic, 2423 reviews, where the median rating was 4.0 (IQR 4) and the mean rating was 3.34. During the pandemic, we processed 2119 reviews with a median rating of 4.0 (IQR 4) and a mean rating of 3.22. Finally, data from the period after the pandemic consisted of 1103 reviews with a median rating of 4.0 (IQR 4) and a mean rating of 3.21. The numeric data can be found in Table 5. These figures were calculated to ensure that there is no noteworthy difference in the numeric ratings of the reviews with text and those without text. As no noteworthy difference was found, we could proceed to the sentiment analysis of the reviews with text.

Finally, we processed the reviews using the AFINN tool (step DA3.1). The results are shown in Figure 7 and Table 6. Figure 6 shows stacked information on these results. The minimum sentiment value obtained using the AFINN tool was −23 (most negative), and the maximum value obtained was 23 (most positive). The median was 0.0, and the mean was 1.01. During the pandemic, the minimum sentiment value was −14, and the maximum value was 15, with a median of 0.0 and a mean of 0.85. After the pandemic, we found a minimum sentiment value of −16 and a maximum value of 24, with a median of 0.0 and a mean of 1.03.

Figure 8 shows the results of the sentiment analysis using the deep learning–based method. Similarly to the lexicon-based method, this method also showed that the distribution of the positive and negative sentiment in Finnish-language reviews stayed relatively stable for the periods before, during, and after the pandemic.

The numerical data obtained are shown in Table 7. As can be observed, there were more positive than negative emotions in the reviews

In the following paragraphs, we present the same analysis for the data from Andalusia as that presented for the data from Finland in the previous section. First, we present the information about the numeric ratings of all reviews (step DA2.1 in Figure 4). The data are shown in Figure 9 and Table 8. As can be observed, contrary to the data for Finland, the ratings changed. The data before the pandemic showed similar ratings as those in the data from Finland, with the median being 4.0 (IQR 4) and the mean being 3.40. During the pandemic, ratings worsened in Andalusia, with the median being 1.0 (IQR 2) and the mean being 1.91. Finally, after the pandemic, the median rating continued to be 1.0 (IQR 3), and the mean rating was 1.99.

After this analysis, we took out all reviews without text and in a language different from Spanish. We show the information for the remaining reviews in Figure 10 and Table 9 (step DA2.2). For the period before the pandemic, we found 7281 reviews, in which the median rating was 3.0 (IQR 3) and the mean rating was 2.88. During the pandemic, negative reviews became more frequent. In this period, we found 13,098 reviews, with a median rating of 1 (IQR 0) and a mean rating of 1.69. Finally, after the pandemic, the data followed the same tendency. We processed 5182 reviews, with a median rating of 1.0 (IQR 1) and a mean rating of 1.91.

As we could expect with these data from Andalusia, the sentiment analysis showed similar results as those for the ratings (step DA3.1). These results are shown in Figure 11 and Table 10. For the period before the pandemic, we processed 8225 reviews, for which we found a minimum sentiment value of −42 and a maximum value of 22. The median was 0.0, and the mean was −0.26. For the period during the pandemic, we retrieved 13,972 reviews, with the minimum sentiment value being −36 and the maximum value being 26.0. The median was −2.0, and the mean was −2.31. For the period after the pandemic, in the 4735 reviews retrieved, the minimum sentiment value was −44.0, and the maximum value was 35.0. The median was −2.0, and the mean was −2.38.

Figure 12 and Table 11 show the results of the sentiment analysis using deep learning–based methods for Andalusia. The results obtained were in line with the previous results obtained using the lexicon-based sentiment analysis tool. The data demonstrate a preponderance of negative emotions over positive emotions. A substantial proportion of the data are attributable to multiclass classification.

Next, we present an analysis of overall word frequencies identified within the Spanish-language (Table 12) and Finnish-language (Table 13) reviews across the 3 defined periods. The tables report the relative frequencies of each word by normalizing the absolute counts by the number of reviews in each period. In addition, we report the occurrences for each of the 10 most frequent words per 1000 reviews for the Spanish (Figure 13) and Finnish (Figure 14) languages.

We can observe in Table 12 that certain words such as “centro” (center), “médico” (physician), “salud” (health), “atender” (to attend or assist), “urgencia” (emergency), “persona” (person), “personal” (staff), and “atención” (attention or care) appeared consistently across all 3 periods, although their frequency fluctuated. A change was observed during the pandemic regarding the frequency of words related to remote communication with the health center.

As shown in Figure 13, the frequency of words related to phone communication—“teléfono” (telephone), “coger” (to pick up), and “llamar” (to call)—increased significantly during the pandemic before declining afterward. In contrast, the frequency of the word “cita” (appointment) not only saw a sharp rise during the pandemic but also continued to increase even further in the postpandemic period. The frequency of the word “vergüenza” (shame) also experienced a notable increase during the pandemic, and although it decreased slightly after the pandemic, it remained much higher than prepandemic levels.

For the Finnish language, we can observe in Table 13 that the most frequent words revealed persistent themes across the periods. Words related to health care delivery, such as “palvelu” (service), “saada” (to get or receive), and “lääkäri” (physician), consistently ranked among the top 3 terms, although their relative order varied.

Interestingly, we found a difference regarding specific socioemotional terms in both regions. In Andalusia, we observed a rise in the use of the word “vergüenza” (shame), entering the top 10 words (position 9) during the pandemic. Conversely, in Finland, the adjective “ystävällinen” (friendly) moved from the eighth most used term to the fourth most used term overall.

Examples of extremely positive and negative Spanish- and Finnish-language reviews translated into English and anonymized are shown next, the Finnish language examples in Table 14, along with the sentiment scores obtained through both the lexicon- and deep learning–based methods.

In the following is an example of an extremely positive review translated from Spanish. The words that appear in the list of the 10 most common words in Table 12 are marked in italics. The lexicon-based sentiment score is 27, the deep learning–based sentiment score is 1, and the review text is: “Doctor [person name 1] is wonderful, polite, kind, caring, generous, and a great medical professional. Without a doubt, one of the very best in all [province name]. Her nurse, [person name 2], is also lovely and caring. I love this health center because you leave with the feeling that you’ve been healed. A ten out of ten score. They are all incredible workers. A pleasure dealing with them.”

In the following is an example of an extremely negative review translated from Spanish. The words that appear in the list of the 10 most common words in Table 12 are marked in italics. The lexicon-based sentiment score is –32, the deep learning–based sentiment score is –1, and the review text is “I’m giving it one star because I can’t give zero. The doctors are not to blame, but the service at the front desk is lamentable. They don’t know how to do their job, they can’t schedule an appointment correctly, they don’t even know how to use the computer program, and on top of that, the ladies treat you with appalling rudeness (or outrageous arrogance). They accept their lack of ability as something normal, and sadly, they have normalized it in their day-to-day work. The waste of time is continuous, and the time they make citizens waste who come to be seen is something that needs looking into. In our specific case, they have messed up an appointment four times for a well-child checkup. First, they got the doctor wrong, then they sent us to the nurse instead of the pediatrician, today the doctor didn’t show up and nobody notified us, and to top it all off, they give me a telephone appointment after I spent 20 minutes demanding a proper appointment...My daughter is 15 months old and hasn’t been seen by a pediatrician for 11 months. These ladies should have to demonstrate their competence for the position they hold, and if they don’t fulfill their duties or don’t know how to perform their job correctly, replacements who are capable of doing those tasks should be sought. It’s a public position that we all pay for, dependent on the Andalusian Health Service, and we citizens who come to be attended to should not have to waste our time like this. Shameful!”

This study showed how a systematic data analysis pipeline can be built for monitoring the perceived quality of primary health care using UGC. The analysis pipeline was tested using Google Maps reviews from primary health care locations in Finland and Andalusia, Spain. Similar approaches have not been previously implemented for Finland or Spain. Google Maps reviews have traditionally been mostly used in assessing perceived service quality in the hospitality industry, but exploiting them in the domain of health care is still in its infancy. There is some previous work in Asian countries such as Taiwan [17] and Malaysia [18] but not in Europe. However, mining user-generated social media data is a cost-effective way of monitoring perceived service quality when compared to surveys [23].

In addition to showing how the data analytics pipeline can be built, we found that the COVID-19 pandemic did remarkably affect the perceived quality of service of primary health care in Andalusia. As shown in Tables 10 and 11, the median values for the sentiment analysis of the textual reviews were neutral before the pandemic. However, both during and after the pandemic, the median of the lexicon-based sentiment analysis results was –2 (IQR 1), and the median of the results of the sentiment analysis conducted using the deep learning–based model was –1. Regarding numeric user ratings for Andalusia, we observed similar changes for worse during and after the pandemic. The mean user ratings before the pandemic were remarkably higher than during and after, as can be observed in Figure 9 and Table 8. The number of reviews with the most positive rating (5) decreased since the pandemic began. This is in line with the findings of Lopez-Picazo et al [52], who presented results on inpatients’ perceived quality based on the hospital regular monitoring plan and the net promoter score in a Spanish hospital. They studied the perceptions before and during the COVID-19 pandemic and reported similar results to those of our study regarding Andalusia—the perceived quality of health care was remarkably lower during and after the pandemic than before it. The lexical analysis revealed important patterns in how health care experiences were described before, during, and after the pandemic. We found an increasing salience of terms related to access and communication as the pandemic progressed. In Spanish-language reviews, nouns such as “teléfono” (telephone) and “cita” (appointment) dominated negative feedback, especially during and after the pandemic, possibly reflecting dissatisfaction with remote communication and appointment systems or availability. Similarly, in Finnish-language reviews, the term “soittaa” (call by phone) appeared on the list of the 10 most frequent words only during the pandemic. For both languages, positive nouns related to service and staff remained highly frequent across periods, suggesting appreciation and acknowledgment.

This paper presents results of a longitudinal study on perceived service quality in health care and the possible effect of a large-scale event (the COVID-19 pandemic) on it. The body of work on this topic is scant, especially when it comes to studies based on sentiment analysis of UGC. The work by Lopez-Picazo et al [52] presents results on the perceived service quality in a Spanish hospital before and during the pandemic, but it is based on surveys. Ours is the first study to address changes in sentiment in UGC regarding health care services before, during, and after the COVID-19 pandemic in Europe. The work by Aduragba et al [53] compares emotions in tweets geolocated in London during March 2019 and March 2020. They observed a rise in the negative emotions of sadness, anxiety, and annoyance when comparing 2020 to 2019. However, their work does not address health care service providers.

Another finding of our study was that the sentiments did not vary in Finland. The mean of both numeric user ratings and sentiment in reviews stayed the same before, during, and after the pandemic. There were very modest differences in the mean values and in minima and maxima. The results concerning the perceived quality of health care in Finland are new and interesting as there is no previous research in that domain. From the publicly available customer satisfaction data collected by Finnish primary care organizers, we know that 77% of the customers completely agree with the following statement: “I received the service/treatment when I needed it” [54]. Moreover, data from Finland show how access to care improved after the outbreak of COVID-19. In October 2019, a total of 40% of nonurgent physician’s appointments in primary health care happened within <7 days. In March 2024, the figure was already almost 60% [55]. In Andalusia, on the contrary, the mean waiting time to access primary health care tripled since December 2018. In 2018, the mean waiting time was 3.8 days, and in 2023, it was 10.4 days [56]. We know that, in Spain and many European countries, the perceived quality of health care decreased during and after the pandemic [52,57], and we know that, according to the study by Tuczyńska et al [57], the perceived quality has remained relatively stable in the United Kingdom. The work by Ainley et al [9] shows results for the United Kingdom along the same lines. However, their work focuses on the delivery of remote care, and the data were analyzed only until October 2020. Tuczyńska et al [57] studied the changes in perceived quality of health care services in 8 European countries. They report a decrease in all the studied countries except the United Kingdom. However, Finland and Andalusia were not addressed in their study.

This study has several limitations. First, we did not control for overall change in ratings and sentiment in Google Maps data for Finland or Andalusia. Thus, we do not know if, for example, the ratings and sentiment for all locations in Andalusia dropped since the beginning of the pandemic. However, previous studies using data from other locations do not show an overall drop in sentiment. A previous study by Aunimo and Martin-Domingo [58] assessed the overall sentiment in UGC on airport services before and during COVID-19 pandemic, and the results showed that the overall sentiment stayed almost the same with a slight increase and only COVID-19–related content showed a decrease in sentiment. The study by Ainley et al [9] analyzed user sentiment toward access to and quality of remote care and noted an increase in positive sentiment in the time after the COVID-19 outbreak in comparison to the time before. However, it would be beneficial if a future study analyzed either the overall sentiment of Google Maps data before, during, and after the COVID-19 pandemic in Andalusia or the sentiment regarding other services such as hotel and travel or public schools to investigate the extent of the drop in sentiment in the area.

Another limitation of our study is that we did not control for possible changes in sentiment due to other simultaneous transformations in the health care systems. For instance, in Finland, there was a major reorganization of the entire public health care system on January 1, 2023. This major change is observed in the last part of the data for the period during the COVID-19 pandemic and in the entire dataset for the period after the COVID-19 pandemic. However, the key primary health care providers (such as health centers or terveyskeskus and health stations or terveysasema) stayed the same, and some differences in their names were compensated by the semantic search mechanism on Google Maps. The collected dataset is consistent for the purpose of this analysis.

A third possible limitation is that, while user ratings and reviews are very popular for some locations such as restaurants and hotels, Google Maps may not be as popular a forum for patients to express their sentiment regarding health care service quality. This may hold true especially for patients in Finland, where the number of reviews in proportion to the population was lower than that for Andalusia. Finland has 5.5 million inhabitants and 12,252 user ratings, whereas Andalusia has 8.5 million inhabitants and 42,796 user ratings. One reason for a lower number of ratings and reviews could be that some important locations such as schools that also provide public primary health care services were left out of the Finnish dataset. Another reason for a lower number of Finnish ratings and reviews may be that the working population typically does not use the public primary health care services as their primary health care is organized through occupational health care services in private health care centers. Thus, demographic characteristics such as age, gender, socioeconomic status, and area of residence may be biased in Google Maps reviews. Moreover, existing studies show that individuals who leave reviews often have more extreme experiences—either highly positive or highly negative—leading to potential bias in the dataset [59]. In addition, older patients (who may be less internet active) are underrepresented [60]. However, user-generated data are not meant to replace official surveys conducted by health care organizations but, rather, are meant to complement the data. Moreover, the aforementioned skews [59,60] do not affect our main research question—sentiment change over time (before, during, and after the COVID-19 pandemic).

This study demonstrated a considerable change in a negative direction in the sentiment of public discussions on primary health care in Andalusia starting from the beginning of the COVID-19 pandemic. However, no change was observed in the sentiment of the corresponding public discussions in Finland. This can be observed in both the numeric user ratings and the sentiment analysis of the review texts. The second finding of this study is that there is a positive correlation between numeric ratings and sentiment analysis of textual reviews, confirming the reliability of the sentiment analysis tools used. Third, we illustrated how Google Maps reviews may be used to monitor public opinion on primary health care. This study provides novel information on the effect of the COVID-19 pandemic on the perceived service quality in primary health care. It also provides insight into how Google Maps reviews may be used to monitor the sentiment of reviews related to hundreds of primary health care locations. Previous research using Google Maps reviews has typically focused on a single location with many reviews, such as a university hospital. Our study presents a methodology for identifying primary health care locations in a specified area and subsequently processing reviews related to them. Further investigations should be conducted to find out why the perceived service quality of primary health care measured using UGC in Andalusia shifted considerably toward negative sentiment since the start of the COVID-19 pandemic and why this was not the case for Finland. For example, analysis of policy response measures reveals that these were stricter in Spain throughout the pandemic (see the COVID-19 stringency index comparison in Figure 2). The impact of other factors such as culture and efficiency of health systems [61] should also be analyzed.