Pandemics have repeatedly caused significant loss of life and triggered severe economic crises worldwide [1]. In 2019, the world encountered one of the deadliest pandemics in modern history, the SARS-CoV-2, also known as COVID-19, which resulted in approximately 18.2 million deaths worldwide from January 1, 2020, to December 31, 2021 [2]. SARS-CoV-2 belongs to the coronavirus family and is primarily transmitted from human to human. Among coronaviruses, it is considered the most contagious [3]. Due to its rapid global spread, the World Health Organization officially declared it a global pandemic on March 11, 2020 [4]. In Sweden, the first COVID-19 case was reported on January 31, 2020, involving a traveler who had arrived from Wuhan, China [5]. By mid-March 2020, the virus had spread across all regions of Sweden, posing significant challenges for public health surveillance and health care systems [6,7]

Effective monitoring of community-level disease transmission is crucial for public health and health care systems [8]. To predict and monitor disease transmission, evaluate the course of the pandemic, and assess the impact of public health intervention policies, epidemiological mathematical models are often used [9]. Such models have proven to be valuable tools for policymakers, enabling data-driven decision-making by exploring different scenarios and their potential outcomes in a controlled manner [10,11]. Among these, the susceptible-infected-recovered (SIR) model is widely regarded as one of the most fundamental and influential frameworks for describing the transmission dynamics of infectious diseases [12]. Over time, several extensions of the SIR model, including susceptible-infected-recovered-susceptible, susceptible-exposed -infected-recovered, and susceptible-infected-susceptible, have been developed to improve short-term infection forecasting [13-15].

While various approaches exist, much of the focus has traditionally been on confirmed cases identified through widespread testing programs. However, traditional case counts are often insufficient for early detection of novel infectious diseases due to limitations in testing capacity and laboratory resources [6]. Syndromic surveillance offers an effective alternative, providing a more comprehensive view of disease dynamics [16-18]. For instance, in the United States, researchers developed a smartphone-based syndromic surveillance system and identified loss of taste and smell as strong predictors of COVID-19 infection [16]. Similarly, app-based self-reported symptoms data have been used to predict COVID-19 cases in Sweden, England, and the United Kingdom [19-23]. Recent studies have indicated that integrating data-driven and epidemiological models, when combined with syndromic and digital data sources, can improve the rapid detection and prediction of infectious diseases. Mathematical and computational models have not only guided immediate responses but also informed medium- and long-term strategies, playing a crucial role in policy planning, public health interventions, and global coordination during the pandemic [24]. Similarly, in a study conducted in 2021, the authors used time series of online search query frequencies to gain insights into COVID-19 prevalence across multiple countries [25]. In these studies [26,27], the authors extended classical compartmental models by incorporating media coverage as a behavioral and informational factor influencing disease transmission dynamics. Drew et al [28] introduced a COVID-19 Symptom Study app in 2020 to collect large-scale, real-time data on symptoms, risk factors, and outcomes, which was subsequently used for scalable epidemiological analysis and modeling to inform public health response.

To respond effectively to outbreak events, policymakers must be well-prepared, relying on real-time and historical data collected from epidemiological, health care, and social processes [29-32]. In Sweden, syndromic surveillance has become a valuable complement to traditional public health monitoring, offering real-time insights into the spread of COVID-19 and other infectious diseases [19]. One important source of syndromic data is Sweden’s National Telehealth Service, the 1177 health care helpline, which often serves as a first point of contact with the health care system for patients. While data are captured for administrative purposes, they have been used in several epidemiological studies for early detection of water-borne disease outbreaks and during the COVID-19 pandemic [33-35]. For example, Spreco et al [34] used aggregated 1177 data as an early signal to estimate potential health care demands in the upcoming weeks. Moreover, a study conducted in 2024 found that calls to Sweden’s 1177 health care helpline reporting COVID-19–related symptoms were important predictors of hospitalizations during the early phase of the pandemic [35].

While previous studies highlight the potential of syndromic surveillance data for disease monitoring, their added value when linked to comprehensive individual-level health register information for improving model accuracy remains poorly understood. In contrast to previous approaches based primarily on aggregated syndromic surveillance data, our study integrates individual-level 1177 symptom data with linked health register information within a compartmental transmission model, allowing individuals to be followed from symptom reporting through subsequent testing and severe disease outcomes. Therefore, in this paper, we extend traditional mathematical models to incorporate a compartment for suspected infection based on individual-level syndromic surveillance data linked with health register data. This approach enables the model to capture early symptomatic signals at the individual level and connect them with subsequent testing outcomes and severe disease progression. The aim of this study is to assess the effect of this integration on the accuracy of short-term infectious disease predictions using a model without syndromic surveillance data as a benchmark.

The number of callers reporting each symptom, together with the corresponding counts of confirmed infections (I_c), hospitalized, and no confirmed infection, was summarized in Table 2. A substantial number of individuals reporting symptoms were later classified as having no confirmed infection (unconfirmed symptomatic callers with respect to COVID-19). In particular, 312,784 (94.78%) individuals reported breathing difficulties, 468,379 (94.41%) reporting fever, and 326,623 (95.43%) reporting cough were identified as having no confirmed infection. This group includes individuals who tested negative as well as those who were not tested and were not subsequently observed in any other compartment within a 14-day follow-up period. This high number of unconfirmed symptomatic callers may reflect heightened public concern, stress, and anxiety during the pandemic, particularly during the initial phase in March-April 2020 when 1177 call volumes increased sharply. In addition, the limited availability of COVID-19 testing during this early phase may have resulted in symptomatic individuals remaining unconfirmed despite potentially having an infection.

The MAE and MAPE of infected confirmed cases (I_c) for the models with and without 1177 data, along with their difference over time, are presented in Figure 2 (panels A and B). The time series analysis shows clear temporal variation in model performance. Early in period 1, the model incorporating 1177 data generally exhibits lower prediction errors compared to the model without 1177. However, during the summer starting period (period 2), both models show increased error, with relatively higher variability in the model with 1177. Late in period 3 (November-December), the model with 1177 again demonstrates improved performance, with lower error compared to the model without 1177. Specifically, the median MAPE difference was 5.26% (IQR –4.80% to 25.25%) in period 1, –0.07% (IQR –22.96% to 46.22%) in period 2, and 0.0008% (IQR –1.75% to 0.07%) in period 3.

Furthermore, box plot analysis (Figure 2, panels C and D) compares the distribution of MAE and MAPE for both models across the 3 periods. In period 1, the model incorporating 1177 shows a median MAPE of 31.37% (IQR 20.03%-39.96%), compared to 33.33% (IQR 25.73%-58.86%) for the model without 1177. In period 2, the model with 1177 yields a median MAPE of 82.90% (IQR 74.51%-160.5%), compared to 98.13% (IQR 63.82%-168.5%) without 1177. In period 3, the model with 1177 reports a median MAPE of 59.35% (IQR 35.72%-64.11%), whereas the model without 1177 shows a median MAPE of 56.94% (IQR 49.35%-64.12%).

Finally, Figure 2 (panel E) illustrates the fitting and prediction performance of both models. The model incorporating 1177 data achieves a MAE of 488.58, whereas the model without 1177 data shows a substantially higher MAE of 16901.38, indicating poor predictive performance in the absence of helpline data.

Figure 3 (panels A and B) presents the MAE and MAPE of hospitalized cases (H) for the models with and without 1177 data, along with their temporal difference. The time-series analysis reveals clear variations in model performance across different epidemic periods. Early in period 1, the model incorporating 1177 data generally performs better. However, during the summer period (period 2), both models exhibit poorer performance, with slightly higher errors observed in the model with 1177. In the subsequent period 3, the relative performance alternates, with the model with 1177 showing improved accuracy at certain time points, while the model without 1177 performs better at others. Specifically, the median MAPE difference was 0.0097% (IQR –2.69% to 8.57%) in period 1, 0.42% (IQR –9.38% to 42.24%) in period 2, and –1.27% (IQR –9.08% to 1.19%) in period 3.

Panels C and D in Figure 3 present box plots of MAE and MAPE for both models, enabling a period -wise comparison. In period 1, the model incorporating 1177 achieves a median MAPE of 11.72% (IQR 9.17%-16.05%), compared to 15.73% (IQR 9.06%-35.95%) for the model without 1177. In period 2, the model with 1177 yields a median MAPE of 57.33% (IQR 35.47%-70.77%), which is lower than the 70.73% (IQR 41.68%-98.95%) observed for the model without 1177. In period 3, the model with 1177 reports a median MAPE of 39.44% (IQR 23.60%-53.76%), whereas the model without 1177 shows a slightly lower median MAPE of 33.80% (IQR 17.69%-48.70%).

Finally, panel E in Figure 3 illustrates the fitting and prediction performance for hospitalized cases. The model incorporating 1177 data achieves a MAE of 63.86, compared to 2128.59 for the model without 1177 data, demonstrating that inclusion of 1177 data substantially improves the model’s ability to capture hospital burden, particularly during the early stages of the pandemic.

Figure 4 (panels A and B) presents the MAE and MAPE of deaths (D) for the models with and without 1177 data, along with their temporal differences. The time-series analysis indicates that both models perform well throughout the evaluation period, with the model incorporating 1177 data showing improved performance during period 1. Overall, the difference in MAPE between the 2 models remained small, with the median MAPE difference remaining below 1.20% across the entire evaluation period, suggesting comparable predictive performance for mortality estimation.

Panels C and D in Figure 4 present box plots of MAE and MAPE distributions across the 3 evaluation periods. In period 1, the model incorporating 1177 data achieves a median MAPE of 1.71% (IQR 1.01%-3.93%), compared to 2.93% (IQR 1.48%-22.50%) for the model without 1177 data, indicating improved accuracy when incorporating symptom-reporting information. In period 2, both models demonstrate similar performance, with the model including 1177 data yielding a slightly lower median MAPE of 0.36% (IQR 0.17%-0.69%) compared with 0.38% (IQR 0.07%-0.74%) for the model without 1177 data. In period 3, the model without 1177 data performs marginally better, with a median MAPE of 0.94% (IQR 0.04%-3.07%) compared with 1.06% (IQR 0.28%-3.30%) for the model incorporating 1177 data.

Finally, panel E in Figure 4 illustrates the fitting and prediction performance for deaths. The model incorporating 1177 data achieves a substantially lower MAE of 177.78 compared with 4184.68 for the model without 1177 data, highlighting the benefit of incorporating 1177 information for improving mortality prediction.

Overall, incorporating 1177 health care helpline data improved forecasting performance across all outcomes, including confirmed cases, hospitalizations, and deaths during period 1 of the pandemic. In period 2, the model incorporating 1177 data continued to show slightly better forecasting performance compared with the model without 1177 data, although the overall performance of both models was less favorable than in period 1. In contrast, during period 3, the model without 1177 data performed better than the model incorporating 1177 data, particularly for confirmed cases and hospitalizations. These findings indicate that the forecasting benefit of incorporating 1177 data was most evident in Period 1, remained present but weaker in period 2, and was not consistently observed during period 3.

To provide a comprehensive assessment of model performance, both median errors and average reductions were evaluated over the evaluation period (May 1, 2020, to December 1, 2020). Median MAPE values showed marginal differences between models with and without 1177 data across all outcomes, being 61.6% (IQR 37.6%-79.6%) and 61.4% (IQR 40.6%-91.1%) for confirmed cases, 37.5% (IQR 18.9%-57.5%) and 41.0% (IQR 19.1%-66.0%) for hospitalizations, and 0.65% (IQR 0.3%-1.9%) and 0.74% (IQR 0.3%-2.2%) for deaths. In contrast, the average reductions in MAPE were 15.7, 8.17, and 1.99 percentage points, while the corresponding average reductions in MAE were 685, 53.40, and 101 for confirmed cases, hospitalizations, and deaths, respectively, indicating the greatest improvement for case predictions and confirming overall improved predictive performance following incorporation of 1177 health care helpline data.

In this study, 2 epidemiological models were developed and evaluated: a baseline SI_cHRD model without 1177 data and an extended SI_sI_cHRD model integrating individual-level syndromic surveillance data from the Swedish 1177 health care helpline. The aim was to assess the contribution of health care helpline data to predicting COVID-19 infections. To contextualize model performance, the calling volume data were divided into 3 temporal periods: period 1 (March 2 to June 10, 2020), during which call volumes increased from 19,785 to 48,869; period 2 (June 11 to August 31, 2020), where call volumes gradually decreased from 48,555 to 22,440; and period 3 (September 1 to December 31, 2020), during which call volumes further declined, ranging from 22,515 to approximately 6000. It was observed that higher call volumes during period 1 coincided with lower prediction errors and more stable model performance. The relatively poorer performance observed during period 2 may be partly attributable to the decline in call volume, which reduced the amount of syndromic information available to the model. In addition, changes in health care-seeking behavior during the summer period, including increased travel and social gatherings, may have altered symptom-reporting patterns. The increased availability of COVID-19 testing services may also have reduced the need for individuals to report their symptoms through the 1177 helpline. During period 3, model performance was less favorable, coinciding with a substantial decline in call volume, as fewer individuals reported COVID-19–related symptoms through the 1177 helpline. In addition, the increasing availability of self-testing toward the end of 2020 and greater public awareness of COVID-19 symptoms may have influenced individuals’ decisions to test and seek health care. Overall, the reduced call volume, together with changes in health care-seeking behavior, testing practices, and epidemic dynamics, may have contributed to the reduced forecasting performance observed during periods 2 and 3. Thus, although 1177 data were associated with improved model performance during the earlier periods of the pandemic, their additional predictive value was less evident toward the end of 2020.

Furthermore, the relatively high median MAPE of 61.6% for confirmed cases indicates substantial error in the point forecasts and limits their interpretation as precise estimates of the absolute number of confirmed cases. Several previous studies have reported substantial variation in forecasting accuracy across models, epidemic phases, and forecasting horizons [37,38]. Therefore, the median MAPE observed in our study should be interpreted in the context of the forecasting horizon and the temporal variation in COVID-19 epidemic dynamics.

Incorporating syndromic surveillance data from the 1177 health care helpline improved short-term forecasting performance, reducing MAPE on average by 15.7 percentage points for confirmed cases, 8.2 percentage points for hospitalizations, and 2.0 percentage points for death cases compared to the model without 1177 data. The model effectively captured early infection dynamics during period 1 and maintained reasonable predictive performance during period 2, whereas its performance declined during period 3. These findings highlight the importance of syndromic surveillance data, particularly during the initial phase of a pandemic.

Syndromic surveillance has previously been used in various contexts, including online search trends, symptom-reporting platforms, and mobile health applications, to support early outbreak detection [28,35]. A study conducted in 2025 found that anomalies detected in 1177 health care helpline calls occurred 7 days before confirmed COVID-19 cases, 13 days before hospitalizations, and 22 days before deaths in Sweden [33]. These findings further support the role of helpline-based syndromic data as an early warning signal for emerging outbreaks. In another study, the authors reviewed syndromic surveillance approaches by examining 4 main data streams: emergency department and hospital records, pharmacy and over-the-counter sales, citizen-generated participatory data, and hybrid multisource systems [39]. Their analysis demonstrated that syndromic indicators could detect outbreaks 2-14 days earlier than conventional laboratory reporting, highlighting their potential for timely epidemic detection.

Previous studies have demonstrated the importance of telehealth-based syndromic data for monitoring infectious disease dynamics. In particular, Spreco et al [34] showed that aggregated 1177 health care helpline symptom data could be used for nowcasting COVID-19 trends. They reported satisfactory nowcasting performance for COVID-19 hospitalizations based on reported adult cough symptoms during the high transmission period (April 5-April 25, 2020), with MAPE ranging from 3%-9%, whereas fever-based nowcasting showed substantially higher errors (14%-52%) during the same period. Their approach relied on analyzing temporal trends in aggregated symptom data. In contrast, our study integrates individual-level 1177 symptom to trace individuals across disease states, enabling direct observation of their trajectories through the different compartments. This allows us to assess what happens following symptom reporting, specifically, whether individuals are subsequently confirmed by testing positive, thereby providing real-world evidence on the predictive value of syndromic data. Notably, during period 1 (May 1 to June 10, 2020), the model incorporating 1177 data achieved a median MAPE of 31.37% for confirmed cases. For hospitalizations, the median MAPE was 11.72%, indicating improved predictive performance compared to confirmed cases. For deaths, the model performed best, with a median MAPE of 1.71% over the same period.

A key strength of our study is that the proposed model can predict confirmed COVID-19 cases, hospitalizations, and deaths during the early stage of the pandemic, when call volumes to the 1177 health care helpline were exceptionally high due to heightened public concern and anxiety. The baseline SI_cHRD model (without 1177 data) showed relatively poor performance during the initial period of the pandemic, likely due to limited testing capacity and reporting delays, which emphasizes the importance of complementary syndromic surveillance data during emerging outbreaks. Incorporating individual-level 1177 symptom data allowed the model to distinguish between laboratory-confirmed cases, unconfirmed symptomatic callers, and hospital cases, thereby reducing predictive errors. These findings highlight that the relative performance of compartmental models and complementary surveillance approaches may vary across different phases of an epidemic or pandemic. During the early phase of an emerging outbreak, when transmission patterns, testing practices, and health care-seeking behavior may change rapidly, syndromic surveillance data may provide useful complementary information when routine surveillance systems are still developing. As the epidemic progresses and surveillance and testing systems become more established, conventional compartmental models may provide useful forecasting information, while the additional contribution of syndromic surveillance may become less evident.

The limitations of this study include reliance on 1177 symptom data, which does not capture individuals who did not contact the service and may therefore miss a substantial number of infections. Although we could follow individuals after symptom reporting using register data, we could not distinguish between untested positive cases and calls related to non-COVID-19–related health issues. Another limitation of this study is that 13 parameters were estimated simultaneously from 3 observed time series. Structural and practical identifiability of the estimated parameters were not formally assessed [40]. Therefore, the uniqueness of the parameter estimates and their reliability under the available data cannot be formally established from the current analysis. Additional factors likely contributing to higher errors include testing delays and underreporting, as well as changes in social behaviors in light of public health recommendations that abruptly altered transmission patterns. Additionally, the use of a daily rolling window may introduce inaccuracies during weekends and holidays. The visualized error trends further highlight the advantage of integrating syndromic surveillance data into infectious disease modeling. Although the baseline SI_cHRD model performed reasonably well during stable transmission periods, its simplified formulation limited its ability to capture complex epidemic dynamics, particularly during periods of rapid trend changes or multiple waves. The fitted curves often followed smooth exponential-like patterns, reflecting the absence of early symptomatic signals and reduced structural flexibility, which made the model less stable during abrupt shifts in transmission.

Future work could focus on expanding the model by integrating additional data streams, such as mobility patterns, demographic characteristics, vaccination information, wastewater surveillance, and internet search trends, to further improve forecasting accuracy while also accounting for time lags and reporting delays [41-43]. In addition, sensitivity analysis and formal assessments of the structural and practical identifiability of the estimated parameters should be conducted to evaluate the robustness and reliability of the model parameter estimates.

Overall, integrating individual-level syndromic data from the 1177 health care helpline improved short-term COVID-19 forecasting performance during the early stage of the pandemic. The extended model captured early changes in transmission and provided more accurate predictions of confirmed cases, hospitalizations, and deaths during this period. These findings highlight the value of incorporating high-resolution syndromic surveillance data into epidemiological modeling frameworks, while emphasizing the need to account for reporting structures and context-specific data limitations. Although this study uses the Swedish 1177 health care helpline, the proposed framework could be adapted to other health care systems with accessible telehealth or syndromic surveillance data. However, its effectiveness and generalizability may depend on differences in health care-seeking behavior, reporting practices, data availability, and telehealth infrastructure across settings.