DEVIATIONS IN INFLUENZA SEASONALITY: ODD COINCIDENCE OR OBSCURE CONSEQUENCE?

Abstract

In temperate regions, influenza typically arrives with the onset of colder weather. Seasonal waves travel over large spaces covering many climatic zones in a relatively short period of time. The precise mechanism for this striking seasonal pattern is still not well understood and the interplay of factors that influence the spread of infection and the emergence of new strains is largely unknown. The study of influenza seasonality has been fraught with problems. One of these is the ever shifting description of illness due to influenza and the use of both the historical definitions and new definitions based on actual isolation of the virus. The compilation of records describing influenza oscillations on a local and global scale is massive, but the value of these data is a function of the definitions used. In this review we argue that both observations of seasonality and deviation from the expected pattern stem from the nature of this disease. Heterogeneity in seasonal patterns may arrive from differences in behavior of specific strains, emergence of a novel strain or cross-protection from previously observed strains. Most likely the seasonal patterns emerge from interactions of individual factors behaving as coupled resonators. We emphasize that both seasonality and deviations from it may merely be a reflection of our inability to disentangle signal from noise, be it due to ambiguity in measurement and/or terminology. We conclude the review with suggestions for new promising and realistic directions with tangible consequences to model complex influenza dynamics in order to effectively control infection.

Introduction

Influenza is a disease of global concern with significant levels of morbidity and mortality, that exhibits both regular seasonal occurrence worldwide and infrequent but devastating pandemics. Research interest in elucidating factors contributing to seasonality is driven both by the desire to understand and explain normal transmission patterns, and also by the conviction that understanding of normal occurrence will provide insight into how outbreaks (local epidemic and/or globally pandemic) occur. This will then allow appropriate resource allocation and will support efforts to mitigate outbreaks. The mechanisms driving influenza seasonality are thought to be related to a number of environmental, agent-specific and host-specific factors. The exact contribution of these factors to seasonality is still largely unknown [1–6]. It is likely that yet undiscovered interplay among host, pathogen, and environmental factors, collectively lead to increased virus transmissibility and infectivity. Great advances have been made in the field of influenza research. Two particular examples are the use of antigenic cartography for vaccine strain selection [7, 8], and the use of advanced whole genome sequence analysis to understand the diversity of the influenza viruses within and across geographically discrete outbreaks [9–11]. The above techniques, along with the ability to perform advanced mathematical modeling informed by surveillance data, provide us with an unprecedented repertoire of tools that can be used for infectious disease forecasting [4, 12–14]. Despite these advances, large lacunae exist in the understanding of the processes that lead to observed seasonal disease dynamics.

In temperate climates, influenza typically arrives with the onset of cold weather, but occasionally breaks out of its expected seasonal pattern. An unusual time of arrival of the last large-scale outbreak of 2009 and a similar occurrence for the pandemic of 1918, triggered an interest in the rules and role of deviations from the expected. A large body of literature comments that while influenza exhibits strong seasonality, the timing, magnitude, and individual characteristics of influenza epidemics change from year to year, place to place, population to population [6, 15–20]. To some extent, heterogeneity in influenza seasonality reflects the nature of the processes of spread, transmission and manifestation of infection. In this communication we argue that our ability to measure and characterize these processes contributes to true and perceived heterogeneity.

In measuring and characterizing influenza seasonality, methodology prone to substantial measurement error produces uncertainty and bias. Research publications and textbooks lack a clear and robust definition of seasonality and the methodology (ranging from determining monthly counts to results of harmonic regression) for assessing seasonality. Further complicating the analysis of seasonal factors is the failure to present findings in a uniform manner and ever-evolving terminology and case definitions for influenza. Finally, investigations into the possibility of multiple mechanisms, aiming to produce observed outcomes has been restricted to at most one or two potential drivers of an admittedly highly complex system. In this review, we focus on three aspects of influenza seasonality, which are critical to any discussion of causal drivers of deviation from seasonal patterns, regardless of the mechanisms from which those seasonal patterns originally emerged: 1) systematic and normalized approaches for depicting disease incidence with existing tools and measurements, 2) the need for a consistent and appropriate terminology related to influenza; and 3) a framework for understanding the full complexity of seasonal oscillations in spatio-temporal dynamics. This framework will help refine and clarify future hypothesis-driven research questions.

Perceived Deviations: heterogeneity of seasonal patterns

We define seasonality in disease occurrence as a temporal pattern of systematic periodic oscillation within a predetermined cycle that can be characterized by peak timing, amplitude and duration (see Figure 1). In general, the cycle might range from months to a few years; however, for simplicity in this communication we refer to an annual cycle, the most common time period. Quantification of a seasonal intensity is based on providing a magnitude of change from a nadir or pre-outbreak level to a seasonal peak. Timing of outbreaks is another important characteristic of seasonality. This concept includes the following aspects of an outbreak time referencing: time of an outbreak onset, time when an outbreak reaches its maximum, time from the onset or from the peak to its end, or a return to a background or pre-outbreak level (see Figure 1). Together with the magnitude these time-related characteristics form a unique outbreak signature. Heterogeneity in seasonality is manifested by variability in peak timing, amplitude and duration. Seasonality can vary by location, population at risk, time period, and the type of health outcome measurement. We postulate that the seasonal oscillation in influenza occurrence is, in fact, a property of a natural process governed by various mechanisms with different ways of manifestation in a given population.

Figure 1.

Seasonal curve and seasonality characteristics. Seasonal curve depicts a temporal pattern of disease occurrence within an annual cycle. Seasonality can be characterized by peak timing, amplitude and duration. Time when a seasonal curve reaches its maximum and duration defined as a time between an outbreak onset and time when a curve returns to a background or level are shown. Seasonal intensity refers to a magnitude of change from a nadir or pre-outbreak level to a seasonal peak.

The variability in these three characteristics is illustrated in a series of maps displaying age-adjusted weekly rates of influenza-related hospitalization at the county level for 4 individual seasons (see Supplement, Movie S1). These dynamic maps depict the spatio-temporal distribution of hospitalizations due to influenza with respect to weekly averages of minimum ambient temperature. The maps illustrate hospitalizations among patients aged 65 and older. These maps demonstrate the emergence of traveling waves of influenza as the ambient temperature drops and allows the visualization of spread or percolation of infection to adjacent areas depending on a spatial distribution of environmental and socio-economic factors relevant to the population. For each presented season, variability was observed in the starting location, duration and the pattern of spread of seasonal outbreaks.

The magnitude of seasonal increase is the most commonly reported parameter of influenza seasonality. In a representative example, the rates of influenza-related hospitalization in the 1999–2000 season were substantially higher compared to other years, as depicted by the intensity of local clusters and global spread. However, within a single season the magnitude may dramatically vary by age even within a seemingly homogeneous age group. For instance, in people aged 65 and older, the oldest category exhibited the highest rate of hospitalization (see Supplemental Figure S1). Considerable heterogeneity in the spread and seasonal magnitude of seasonal influenza has been documented in populations of high vulnerability (i.e., at high risk for exposure to influenza or severe health outcome) [21, 22]. While the magnitude of seasonal increase in influenza morbidity and mortality in a specific population might reflect the behavior in the general population, a simple comparison of seasonal magnitudes across the various groups has to be attempted with caution. Furthermore, inferences from observed differences in the seasonal intensities have to be considered carefully due to potential diversity in causal underlying mechanisms. For example, the risk factors implicated to spatiotemporal patterns for children, young adults, or adults with young children, might have low or even no relevance to the patterns of hospitalization due to influenza among the elderly. The same logic applies to the comparison of seasonal peaks across regions and locations. Approaches to quantification of the seasonal magnitude also vary, ranging from providing the highest value of an outcome of interest (incidence rates, percent of positive tests, number of cases, etc.) observed over a time period of incidence to an estimated trough-to-peak ratio [23]. Measures of magnitude include excess mortality values [4], relative and absolute intensity [20], or their proxies.

Characterization and reporting of a peak timing in influenza are improving. Not only is the month with the highest number of cases consistently provided in literature but also information on the range in timing of the regional peaks [24]. It has been shown that the seasonal peak timing [20] and also the time taken to reach the peak and baseline levels [24] can vary locally and regionally. It has been well documented that Influenza A often precedes Influenza B [25], indicating potential heterogeneity in seasonal peak timing associated with strain diversity. Peak timing and seasonal magnitude can be correlated: earlier outbreaks have higher intensity [20], which can be linked to antigenic drift [26]. The dynamic maps provide insight on the time elapsing from the onset to complete dissolving of a seasonal outbreak, which may take 4 to 7 weeks [20]. The geo-referenced sequence of seasonal peaks forms travelling waves of influenza, and allow us to characterize a global pattern of transmission [27–30].

It has been shown that there is a remarkable degree of synchronization of influenza outbreaks at a regional level [20, 31–33] as well as between countries [24, 34] in temperate climates. Investigators relate synchronization to globalization [35, 36], social mixing patterns [37], and transportation networks [30, 38]. However, it is noteworthy that synchronization in the tropics is not extensively documented in spite of high population density and high-connectivity between regions. In larger countries like China, Brazil and India, certain degree of synchronization is seen for regions that have similar climatic conditions [15, 39, 40].

Synchronization of influenza seasonality with environmental parameters offers promising potential for an integrated forecast of infection on a local and global scale. The link of influenza with low ambient temperature favoring survival of aerosol viruses [41] and indoor crowding [42] might be implicated in a southward trend in occurrence of increased hospitalizations early in the season (September – October) in the Midwest and South. Another easily distinguishable wave of outbreaks from northwest to southeast, corresponding to decrease in temperature, was observed in Texas and Oklahoma in late October through November with a frequent appearance of clusters at temperature gradient fronts. In both coastal regions Atlantic and Pacific, influenza hospitalizations on average peaked almost six days later than in the Central region (between −80° and −100° longitude) at 4.9 weeks versus 5.6 weeks [20]. This suggests that, in general, on a large spatial scale, traveling waves of influenza move from southwest to northeast. However, as illustrated by the dynamic maps, even within a single influenza season it is possible to trace multiple origins contributing to the overall seasonal curve.

A seasonal pattern observed globally is not necessarily a simple sum of patterns observed locally [43]. Although, annual epidemics typically begin abruptly, peak within two to three weeks, and last from five to ten weeks in the continental USA [20], their local behavior might exhibit unusual clustering percolating during an influenza season. In the presented example one of the most striking observations is the presence of clusters of high influenza incidence that occurred early in the fall of 1999, were maintained throughout the influenza season, and were among the last remaining at the end of the influenza season (see Supplemental Figure S1). A potential answer for isolated percolations and for a global seasonal pattern is likely to be related to characteristics of circulating strains which have been dominated in the past and were reoccurring in a given season [44].

A depiction of a typical seasonal pattern or a departure from it requires a spatially-explicit time series modeling, which usually entails the selection or specification of a time period and geographical area. In locations with relatively small populations, an aggregation of data into “meaningfully large” numbers leads to reporting of monthly or even quarterly cases of influenza that severely weakens the quality of analysis. The use of fine time units – days and weeks – allows the detection of seasonal patterns with high resolution; however, it often requires an aggregation over a large, often heterogeneous, geographical area and may conceal an isolated pattern. A departure from what is “typically” observed on a large geographical scale needs to be better characterized with respect to local diversity of circulating strains and the criteria for a proper comparison should be grounded on what we can measure reliably with a sufficient degree of reproducibility, precision, and accuracy suitable for intended purposes and goals.

Influenza: the new Tower of Babel or the sources of Obscured Observations

The paradigm of seasonality and the heterogeneity in patterns observed can originate from the process itself or from our ability to detect and measure seasonality. The lack of sound science-based definitions, reliable data and limited methods for presenting data and assessing statistical significance in temporal oscillations can obscure the true seasonal pattern Influenza, is an ancient term and, ironically, lacked a firm meaning from the start, becoming even more convoluted as time progressed. The term has long been in clinical and public health use, and predated the discovery of the true cause of the disease in 1933. Influenza has also long been used as a blanket term to refer and be synonymous with respiratory illness - conflation with common cold. In both epidemiology and medical realms, the term “influenza” is used restrictively in some cases, referring only to the disease caused by the influenza virus confirmed by laboratory tests [45]. At the other end of the precision scale, influenza refers to a collection of signs and symptoms, which are themselves not clearly defined - and maybe cannot be defined to any degree of clarity [46–50]. The routine use of such wide range of case definitions leads undoubtedly to substantial noise in observed temporal patterns, may produce false alarms, or fail to recognize an unusual departure from a seasonal curve.

Based on the clinical progression of the disease we expect that mild cases will appear first in a community, with the following rise of outpatient visits, an increase in hospitalizations and then death cases according to a pyramidal structure (Figure 2). However, rapid onset of influenza, high infectivity, and heterogeneous herd immunity [51] might obscure this temporal pattern. Furthermore, an event of high media attention might distort otherwise a smooth seasonal curve by a disproportionally high sudden rise or spike in tracked records if the case definition is prone to such fluctuations. With the increase in digital tools for tracking influenza cases over the internet and social media, vague definitions of influenza are currently the core of temporal trends [52]. These new technologies pursue a noble goal to provide an early warning for influenza arrival and their credibility depends on the quality of tracked responses and ability to separate signal from noise. Considering ongoing attempts to actively use text mining in large volumes of medical records, the signal-to-noise-ratio is the most important consideration in understanding the departure from an expected pattern. Until definitions and the approaches to consider various terms in data mining and text search engines are clarified, the temporal oscillations produced by massive text mining algorithms might be severely obscured and the similarities or differences in the detected patterns could be purely coincidental.

Figure 2.

Pyramidal structure of disease burden with respect to severity. As severity of influenza infection progresses from asymptomatic to mild to severe, the number of cases decreases proportionally (Panel A). In a population with high herd immunity the majority of asymptomatic cases may appear unnoticed (Panel B). In vulnerable populations, the proportion of patients with severe outcome might be very large even with a relatively small contribution to overall burden (Paned C). Based on a clinical progression of influenza we expect that outpatient visits for mild cases peak first, severe cases requiring hospitalization and specialized medical care peak next, and cases resulted in death peak last (Panel D).

“Influenza-like illness” (ILI) is currently recognized as the cornerstone of syndromic surveillance and is most often used to refer to persons with signs and/or symptoms that are commonly the result of influenza virus infection. Comparisons of seasonal patterns derived from syndromic surveillance should be taken with caution, since the case definition may change over time and vary from country to country and season by season as the set of symptoms may change. A reliable surveillance system that produces systematically evaluated laboratory-confirmed cases with reasonable spatial granularity and sufficient level of detail on demographic composition and molecular characterization is a key for comprehensive depiction of influenza seasonality. Geo-referenced data uniformly collected and updated on a weekly basis can serve as an indicator of the level of influenza activity for the whole country. The prime example of an established influenza monitoring comes from a number of national systems where elaborate sentinel surveillance is combined with extensive laboratory characterization [33, 34, 53, 54, 55]. The establishment of surveillance system in countries with tropical climates enables the depiction of seasonal trends in locations where historically data were very limited. One example of a tropical city having a good surveillance system is Hong Kong, where the burden of hospitalization can be compared with that of the US and a distinct pattern of seasonality exists [44, 56, 57]. Compilation, validation and retention of certain minimum demographic (e.g. age, gender, location) and clinical information (e.g. disease severity, outcome) in publicly reported surveillance data are likely to increase the utilization and usefulness of monitoring efforts in assessing influenza seasonality.

Hospitalization and medical claims records offer a unique systematic approach to depicting seasonal patterns, which is likely to be different from what is observed via surveillance due to a shift to a population that is likely to be more susceptible or disease prone, such as children, elderly, and people with underlying medical conditions (see Supplemental Figure S2). While weekly pneumonia and influenza-associated hospitalizations has been used as a reliable indicator of the influenza morbidity, it is likely that seasonality, specifically peak timing, depicted by the hospitalization claims contains substantial delays. Imprecision of clinical diagnosis may add to the noise in this seasonal pattern.

Because of the various notions associated with influenza, the attempt to be more precise has gone in the direction of adding more words to qualify the terms. The types of restrictions added include geographical designations, designations related to time of the year, and geography, in addition to an assortment of other types of limits added to the central term. It is useful to disentangle various adjectival categories, and asks whether the added words provides greater precision or merely complicates the fuzziness of the language. The “pandemic flu” is often intended to mean a “highly virulent” that leaves in its wake an excess in mortality throughout the world, whereas “epidemic influenza” may denote a more localized viral infection with perhaps a lower number of deaths. Some epidemics carry an adjective that identifies some aspect of the virus, such as “swine” or “avian” flu, based on the animal reservoir the strains may have originated. Are these simplifications intended for lay audience or merely a reflection of a sloppy language use? An appearance of a term “seasonal” attempts to loosely specify an influenza season with characteristics which are somewhat expected, or at least are not “epidemic” or “pandemic.” Does seasonal influenza stem from seasonal requests for testing, at least in part? Is un-seasonal flu a departure from a due course and should serve as an alarm? These questions need to be clarified.

Coupled resonators to study deviations of seasonality

Many researchers have proposed isolated potential mechanistic drivers of seasonal/periodic fluctuations in influenza [58, 59]. Such studies investigate local patterns within years independent of a broader temporal context or focus on long-term patterns in which variation among individual years is averaged in favor of understanding emerging trends. There is, however, a different type of hypothesis to describe the mechanism by which “deviations” from expected oscillations might arise: coupled resonators. Building on ideas from physics initially proposed in the 1600's, this hypothesis proposes an idea that many different mechanisms may each contribute an oscillatory driver of influenza dynamics, but that the differences in strength, timing, and the potential amplifying and damping effects they have on each other may lead to quasi-chaotic local behavior in an otherwise globally periodic system. (In the language of modern physics, this involves the study of coherence and resonance in loosely-coupled oscillators [60]). Engineers and physicists have already developed incredibly useful theory to describe necessary and sufficient conditions for coherence and resonance behaviors in such systems, including incorporation of the impact of stochasticity and time-delay, making their results not only relevant, but directly analogous to proposed drivers of disease dynamics [61]. This idea is not entirely new to the study of seasonal influenza, but has thus far been confined to studying multiple effects of single, or small sets of, mechanistic drivers of oscillation [23]. These insights have been extremely valuable, but have not yet realized their full potential as a unifying principle from which plural-mechanistic hypotheses may be considered.

Importantly, we do not mean to suggest that the correct choice of action for current research would be to compose a model of “everything but the kitchen sink”, tuning the interactions until observed patterns that include global periodicity with deviations of the observed type emerge. While that would be possible, it would be practically meaningless. We believe the focus of these efforts should shift away from trying to demonstrate which mechanisms may be strong enough to be primary drivers of global patterns, and instead begin to focus on how different mechanisms affect each other (starting with pair-wise interactions, but then also explicitly scaling up experimentally to discover potential three-way and higher-order interactions). Only after this empirical groundwork has begun can models begin to explore at what level these interactions may be appropriate for inclusion into an all-explaining paradigm of seasonality. From this perspective, observations which are currently believed to be deviations from seasonal patterns may actually be the result of a sufficiently complex system of loosely coupled oscillators that, in fact, reveal such seeming anomalies as logical necessities in the global pattern of disease incidence.

Conclusions

In temperate climates, seasonal influenza arrives in late fall to early winter and dissipates in spring. In tropics, annual fluctuations are more complex and linked to water content in the air, rather than ambient temperature cycle. On a relatively global scale, annual epidemics begin abruptly, peak within two to three weeks, and last from five to ten weeks. Seasonal waves travel over large spaces covering many climatic zones in a relatively short period of time. The precise mechanisms governing the peak timing, amplitude, shape, and duration of seasonal waves are unknown. The relationships between host susceptibility, the new strain emergence and their genetic variability, factors that influence the spread of infections and characteristics of seasonality are not well understood. This area of research is extremely promising and already demonstrating a strong potential. To ensure success, we need to shape the meanings, referents of terms, and develop the models to correspond as closely as possible with what we know - and equally important, with what we do not know. Furthermore, for monitoring purposes and determination of endemic levels of disease so that we can accurately read the warning signs in nature, the use of precise terms and the development of novel creative approaches for depicting seasonal patterns and departure from the expected are critical directions of research. The development of rigorous definitions uniformly implemented will form the understanding whether our observations are, in fact, measurements of a biologically fluctuating system, or actually logical necessities of a steady, but stochastic, natural state.