Functional data analysis characterizes the shapes of the first COVID-19 epidemic wave in Italy

Tobia Boschi, Jacopo Di Iorio, Lorenzo Testa, Marzia A. Cremona, Francesca Chiaromonte

TL;DR

This paper uses Functional Data Analysis (FDA) to reveal two distinct COVID-19 epidemic patterns in Italy: an aggressive "exponential" wave in the north and a milder "flattened" wave elsewhere. The study demonstrates that mobility and positivity rates are significant predictors of mortality, even when accounting for other factors, and suggests primary care might mitigate outcomes.

What Problem Does This Paper Address?

The COVID-19 pandemic presented a significant challenge in understanding its spread and impact, particularly due to the heterogeneity observed across different regions. While traditional epidemiological models exist, this paper aims to provide a novel perspective by analyzing the shapes of epidemic curves rather than just peak values or total counts. Specifically, it addresses the question of how regional differences in Italy manifested during the first wave and what factors, beyond simple case numbers, contributed to these variations in mortality. The authors highlight the limitations of public data (accuracy, resolution) and seek to extract meaningful insights despite these challenges.

Methods

The core of this research lies in the application of Functional Data Analysis (FDA) techniques. Instead of treating data points as independent, FDA treats entire curves (e.g., daily mortality over time) as single observations. Key methods employed include:

  • Probabilistic K-mean with local alignment (probKMA): This algorithm groups similar shapes of mortality curves, allowing for shifts in time to align them. This is crucial for identifying recurring epidemic patterns, or "motifs."
  • Interval-Wise Testing (IWTomics): Used to statistically compare the identified epidemic groups and pinpoint the time intervals where they significantly differ.
  • Functional Boxplots and Depth Analyses: Tools for visualizing and ranking the curves based on their centrality and identifying outliers.
  • Function-on-Function Regression: A sophisticated regression technique that models the relationship between entire curves (e.g., mortality curve vs. mobility curve). This allows for the estimation of how changes in one curve over time affect another curve over time.
  • SsNAL-EN: A functional generalization of Elastic Net for feature selection, used to identify the most important scalar covariates influencing mortality.

The novelty lies in applying these FDA tools to epidemiological data, moving beyond traditional statistical models to analyze the temporal dynamics and shapes of the epidemic.

Datasets

The study utilizes several publicly available datasets at the regional level in Italy:

  • COVID-19 Death Counts: Daily cumulative data from the Italian Civil Protection Agency (DPC).
  • All-Cause Mortality Data: Daily mortality rates from the Italian National Statistical Institute (ISTAT) for the years 2015-2019, used to estimate baseline mortality and account for undercounting in official COVID-19 records.
  • Google Mobility Report: Daily data for the "Grocery & pharmacy" category, reflecting local mobility trends.
  • Positivity Rates: Daily ratios of new positive COVID-19 cases to the number of tests performed, derived from Civil Protection Agency data.
  • Socio-demographic, Infrastructural, and Environmental Covariates: A curated set of 12 variables (e.g., % over 65, % diabetics, ICU beds per capita, pollution levels) from various public sources (ISTAT, Ministry of Health, etc.).

Key Insights and Ideas

  • Two Distinct Epidemics: The most striking insight is the clear division of Italy into two epidemic patterns: a rapid, "exponential" rise and fall in the north (especially Lombardia) and a slower, "flattened" curve in the center and south, with Veneto being a notable exception that managed to flatten its curve early.
  • Mobility as a Key Predictor: Local mobility, particularly in the "Grocery & pharmacy" category, emerged as a strong predictor of mortality. Early and mid-epidemic mobility levels were positively associated with peak mortality, while later, increased mobility in milder epidemic regions showed a negative association.
  • Positivity as a Sentinel Indicator: Positivity rates also showed a predictive relationship with mortality, though weaker than mobility, suggesting it acts as a sentinel indicator.
  • Primary Care's Potential Role: Among the socio-demographic and infrastructural factors, the availability of primary care (proxied by adults per family doctor) appeared to mitigate mortality, while contacts in hospitals, schools, and workplaces seemed to aggravate it.
  • Data Limitations vs. FDA Power: The study effectively demonstrates how FDA can extract significant trends and insights even from data with known limitations in accuracy and resolution, highlighting the power of analyzing curve shapes.

Key Results

  • ProbKMA identified two distinct epidemic groups: Group 1 ("exponential") and Group 2 ("flattened").
  • Function-on-function regression showed significant associations:
    • Local mobility and positivity were significant predictors of mortality, explaining a substantial portion of the variability (in-sample R² ≈ 90%, LOO-CV R² ≈ 52% for the joint model).
  • Partial R² values indicated comparable contributions from mobility (≈ 62%) and positivity (≈ 53%) in the MAX data set.
  • Marginal regressions revealed:
    • Primary care (adults per family doctor) was the strongest marginal predictor, showing a large positive association with mortality.
  • Hospital, school, and workplace contagion hub proxies also showed strong positive associations.
  • When controlling for covariates: The strong predictive power of mobility and positivity remained largely intact, suggesting their independent roles.
  • Lombardia's mortality was consistently underestimated by the models, suggesting uncaptured factors or data issues specific to the hardest-hit region.

Limitations and Future Work

The authors are upfront about several limitations:

  • Data Resolution and Accuracy: The regional resolution of the data can dilute signals from finer-scale variations, and official death counts were found to be significantly undercounted, especially early on.
  • Lag Times: Precisely quantifying the lag between mobility/positivity and mortality was challenging due to data limitations.
  • Causality: While strong associations are shown, drawing definitive causal implications is difficult with observational data.

Future work suggested includes:

  • Richer Data: Utilizing more granular data on cases, hospitalizations, and deaths, at a finer spatial resolution (e.g., municipalities), would allow for more nuanced analyses.
  • Extended Temporal Span: Including data beyond April 2020 would help characterize later phases of the epidemic and the impact of policy changes.
  • FDA Tool Development: Further development and validation of FDA techniques for epidemiological analysis.
  • Investigating Specific Factors: Deeper exploration of the roles of chronic conditions, household structures, and ICU availability.

Summary updated

Abstract

We investigate patterns of COVID-19 mortality across 20 Italian regions and their association with mobility, positivity, and socio-demographic, infrastructural and environmental covariates. Notwithstanding limitations in accuracy and resolution of the data available from public sources, we pinpoint significant trends exploiting information in curves and shapes with Functional Data Analysis techniques. These depict two starkly different epidemics; an "exponential" one unfolding in Lombardia and the worst hit areas of the north, and a milder, "flat(tened)" one in the rest of the country-including Veneto, where cases appeared concurrently with Lombardia but aggressive testing was implemented early on. We find that mobility and positivity can predict COVID-19 mortality, also when controlling for relevant covariates. Among the latter, primary care appears to mitigate mortality, and contacts in hospitals, schools and workplaces to aggravate it. The techniques we describe could capture additional and potentially sharper signals if applied to richer data.
BibTeX
@inproceedings{boschi2021functional,
  author = {Tobia Boschi and Jacopo Di Iorio and Lorenzo Testa and Marzia A. Cremona and Francesca Chiaromonte},
  title = {{Functional data analysis characterizes the shapes of the first COVID-19 epidemic wave in Italy}},
  booktitle = {Scientific},
  year = {2021},
  doi = {10.1038/s41598-021-95866-y},
}
Figure