We used cITS to assess the impact of COVID-19 containment interventions on inpatient hospitalizations, average length of hospital stay, and in-hospital mortality following the work of Bernal and colleagues [19]. By analyzing data collected at regular intervals over time, we account for underlying trends when making pre-post comparisons providing more accuracy in our results [19].

We used de-identified inpatient admissions data, inclusive of 41 public hospitals and 10 private facilities across Tuscany, made available quarterly by the Tuscan region. The data were received in September 2020 and analyzed from October 2020 to March 2021. We utilized data from the first week of January to the last week of July during the years 2015–2020. Census data was obtained from the Italian Statistics Bureau, with a projection for 2020 based on previous year trends. The study was carried out in compliance with Italian law on privacy, and approval by an Ethics Committee was not required.

Tuscany has 41 public and 10 private hospitals located within 34 health districts serving 3.7 million citizens. During the first wave of COVID-19 (March 2020), Tuscany was moderately affected compared to Lombardy and Veneto, but had also increased the number of ICU beds to meet the needs of increased demand due to COVID-19 patients. Additionally, in April 2020, telehealth services were activated following a regional act to support the ongoing management and care of non-COVID-19 patients.

Our primary outcome of interest was the weekly number of patients in the hospital per 100,000 population, overall and for specific conditions of AMI, stroke, and heart failure. Our secondary outcomes were the weekly average length of stay, and percent in-hospital mortality, the latter defined as the percentage of existing hospitalizations records with a discharge code of “deceased” out of all hospitalizations per week. While average length of stay is typically used as an indicator of efficiency, we used it as a proxy for organizational changes in healthcare resource utilization. We hypothesized that average length of stay for patients without COVID-19 decreased due to diversion of healthcare resources towards COVID-19 care. To better understand illness severity, we examined the percent of in-hospital mortality. We hypothesized that the percent of in-hospital mortality for patients without COVID-19 increased due to less severely ill patients disproportionately avoiding hospital care during the pandemic.

Our analytic sample consisted of all inpatient hospitalizations between the months of January to July, for the years 2015–2020, excluding COVID-hospitalizations using the International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9, details in Supplementary Material S3). De-identified data were retrospectively retrieved from the inpatient administrative databases, and patients with AMI, stroke and heart failure were identified using ICD-9 codes (Supplementary Material S3). Inpatient hospitalization was categorized as unplanned, planned, or other. Unplanned care hospitalizations consisted of admissions from the emergency department or direct admissions via ambulance arrival. Planned hospitalizations included all pre-scheduled hospitalizations. Other admissions include birthing and psychiatric hospitalizations.

While not required in interrupted time series analysis, we added a control group to account for seasonal effects and thus strengthen causal inference of our results. We selected a historical control as the pandemic’s far reaching impact did not allow for another suitable control group [19]. In this cITS, we defined the exposure group as year 2020 and control as the average of 2015–2019. Historical cohort controls have been previously used to evaluate the impact of a drug funding restriction policy on outcomes such as drug expenditure, primary care visits, and admissions to emergency departments [21]. Our unit of observation was at the patient hospitalization level and our unit of analysis was aggregated at the weekly level. We defined the “interruption” as the combined effects of the COVID containment measures and fear of infection since it is impossible to disentangle the specific contributions of each. We defined week 10 of 2020 as the time of “interruption”; that is when the Italian government issued a nationwide lockdown. We split the data into three separate time periods—pre-lockdown, phase-in, and post-lockdown (Table 1). With the first COVID-death registered on 22nd February 2020 (week 8) and the announcement of a nationwide lockdown on 10th March 2020 (week 10), we expected that the combined influence of increased media coverage and local measures to affect healthcare seeking behaviour gradually initially, and abruptly following the nation-wide lockdown [22]. Thus, we specified weeks 8–10 as the phase-in period and were excluded from analyses (insufficient time points to specify as a separate time period) [23].

We used segmented linear regression with an autoregressive error model to measure the size of the intervention’s immediate effect (level change), effect on changes in the slope (trend change), and an estimate of the longer-term effect [23–26] (for further details see Supplementary Material S4). Autocorrelation was assessed using Durbin-Watson tests, and the autocorrelation function and partial autocorrelation function were used to identify lag order for model correction. We fit a single interrupted time series as an ordinary least squares model and used autocorrelation function and partial autocorrelation function to identify the lag order and test for autocorrelation in the model error distribution. We then fitted a cITS, model specified indicated here: y = β 0 + β 1 ⋅ w e e k + β 2 ⋅ exp o s e d + β 3 ⋅ w e e k ⋅ exp ⁡ o s e d + β 4 ⋅ p o s t + β 5 ⋅ w e e k ⋅ p o s t + β 6 ⋅ p o s t ⋅ exp o s e d + β 7 ⋅ w e e k ⋅ exp o s e d ⋅ p o s t + ε

To ensure that the parallel trend assumption for cITS hold and that comparability between treatment (2020) and control (average of 2015–2019) on pre-intervention covariates exist, we confirmed that the pre-intervention trend differences (β₃) were not significant (p > 0.05) [27]. In models where the pre-intervention trend differed between treatment and control, we opted for the most recent year where the trend difference was not significant to serve as control. We report the level change, trend change, and a long-term estimate of the interruption effect for each interrupted time series model. Level change (β₆) is the difference in mean scores before and after the interruption and represents the size of the interruption’s immediate impact. To facilitate easier interpretation, we also report level change in relative terms (termed relative change thereafter) calculated as the percentage change relative to the counterfactual. Trend change (β₇) represents the change in slope gradient following the interruption and quantifies the interruption effect on the overall mean. And lastly, we report the interruption’s estimated longer-term effect in the last week of August using the projected level changed based on the modeled counterfactual trend extended into the post-interruption period.

We conducted subgroup analyses by stratifying the population by sex (male and female), and age groups (under 60, 60 to 69, 70 to 79, 80 to 85, and over 86).

All analyses were conducted in R 4.0.3. All p-values are 2-sided and a value of p < 0.05 was considered statistically significant.

There were 255,882 non-COVID inpatient hospitalizations involving 206,115 patients (54% female; mean age of 57.2 ± 26.9 years) during January-July 2020 (Table 2). The inpatient hospitalization rate decreased by 182/100,000 (95%CI: -235, -130) following the interruption, representing a 56% decrease (Figure 1A; Table 3). This level change was followed by an insignificant trend increase of 1.04 per 100,000 population per week (95% CI: −8.16, 10.24).

Subgroup analyses showed no significant differences in inpatient hospitalization rates between male (−88, 95% CI [−116, −60]) and female (−89, 95% CI [−117, −61]). We did find decreased hospitalization rates of patients under the age of 60 (−87, 95% CI [−115, −59]) compared to other age groups (Supplementary Material S7A).

Unplanned hospitalizations represented 46.5% of all non-COVID inpatient hospitalizations, which is higher than the average of 40% seen in previous years. Unplanned hospitalization remained relatively stable at 116 per 100,000 population prior to the interruption. This was followed by an immediate decrease of 39/100,000 (95% CI: −51.1, −26.0) hospitalizations, representing a 32% decrease (Figure 1B; Table 3). A non-significant positive trend of 0.65 per 100,000 population per week (95% CI: −1.99, 3.34) was observed post interruption, with a rate of 92/100,000 population in the last week of July.

Subgroup analyses show no significant differences in unplanned care hospitalization rates between male (−21, 95% CI [−27, −14]) and female (−22, 95% CI [−29, −15]). Hospitalization rates decrease was larger for those under the age of 60 (−15, 95% CI [−20, −9]) compared to other age groups (Supplementary Material S7B).

In 2020, 1,333 patients (34% female; mean age of 72.7 ± 13.0 years) were hospitalized for AMI at a markedly lower rate than in previous years (4.06 compared to 5.40/100,000, Supplementary Material S6A). Following the interruption, there was a level change of −1.32/100,000 (95% CI: −1.98, −0.66) in AMI hospitalizations, representing a 30% decrease compared to counterfactual (Table 3). This was followed by an insignificant trend change of (−0.04, 95% CI: −0.2, 0.1). By the last week of July 2020, AMI hospitalizations had recovered somewhat to 3.36/100,000. Average length of stay for AMI hospitalizations was stable pre-interruption with no significant level change following interruption (0.6 days, 95% CI: −0.31, 1.57, Supplementary Material S6B). There was no significant in-hospital mortality changes for AMI hospitalizations (−1.39, 95% CI: −4.73, 1.94).

Subgroup analyses show no significant differences in hospitalization rates between male (−0.84, 95% CI [−1.62, −0.06]) and female (−0.55, 95% CI [−0.80, −0.31]) (Supplementary Material S7C). There were no obvious differences in hospitalization rates between age groups. There was a non-significant larger decrease in hospitalization for NSTEMI (−0.84, 95% CI: −1.39, −0.28; 32% reduction compared to counterfactual) than for STEMI (−0.59, 95% CI: −1.06, −0.11; 12.6% reduction compared to counterfactual; Supplementary Material S7C).

In 2020, 5,317 patients (49% female; mean age of 76.3 ± 13.7 years) were hospitalized for stroke, which is comparable to previous years (6.30 compared to 6.23/100,000, Supplementary Material S6D). Following the interruption, a level change of -1.51/100,000 hospitalizations (95% CI: −2.57, −0.44) occurred, representing a 23% decrease in hospitalizations, with no significant trend change (0.05/100,000, 95%CI: −0.13, 0.24) (Table 3). By the last of July 2020, stroke hospitalization rate was at 4.92/100,000. The average length of stay remained quite stable during the months between January and July in 2020 with a small, insignificant increase of 0.82 days (95% CI: −0.94, 2.58) after the interruption. In-hospital mortality rate for stroke hospitalizations was lower in the months of January and February of 2020 compared to the average of previous years. We observed an increase of 5% (95% CI: −0.27%, 10.82%) after the interruption, bringing the in-hospital mortality rate to 14%, which was comparable to previous year.

Subgroup analyses show no significant differences in hospitalization rates between male (−0.86, 95% CI [−1.45, −0.27]) and female (−0.75, 95% CI [−1.41, −0.09]) (Supplementary Material S7D). We did not observe significant differences in hospitalization rates between patients admitted with ischemic (−0.92, 95% CI [−1.60, −0.26]) and hemorrhagic stroke (−0.69, 95% CI [−1.26, −0.11]). We did find seniors aged 80 to 85 and 86 years and over to have a larger reduction in hospitalization compared to the younger age groups (80–85: −0.9, 95% CI (−1.2, −0.5); over 86: −0.7, 95% CI (−1.2, −0.3)).

In 2020, 12,263 patients (49% female; mean age 81.2 ± 11.3 years) were hospitalized for heart failure, which was comparable to previous years (17.45 compared to 18.20/100,000, Supplementary Material S6G). Following the interruption, there was a level change of −8.71/100,000 hospitalization (95% CI: −11,13, −6.29), representing a 45% reduction compared to counterfactual (Table 3). The impact of this level change was sustained, with no significant trend changes (0.33/100,000 hospitalizations, 95% CI: −0.09, 0.75). The average length of stay remained quite stable between the months January to July in 2020 (level change: 0.01 days, 95% CI (−0.69, 0.71); trend change: −0.06, 95% CI (−0.19, 0.06)). Similarly, no significant changes were observed in percentage of in-hospital mortality during the study period (level change: 2.43%, 95% CI (−0.73, 5.60); trend change: −0.36, 95% CI (−0.91, 0.19), Table 3).

Subgroup analyses show no significant differences in hospitalization rates between male (−4.55, 95% CI [−5.96, −3.14]) and female (−4.65, 95% CI [−6.14, −3.16]) (Supplementary Material S7E). We observed a larger decrease in hospitalization among seniors over the age of 86 (−3.11, 95% CI [−3.96, −2.26]) compared to other younger age groups.

Our study has important limitations. First, due to the rapidly evolving and unpredictable nature of the pandemic, we were unable to pinpoint a specific time point when the interruption took effect. Additionally, we were unable to differentiate between more direct impacts of the pandemic and the associated containment measures. To account for both, we had specified a two-week phase-in period to such that both events to take effect. Second, we did not have access to mortality data outside of the hospital, and thus were unable to determine the full extent of impact in delayed care for time-sensitive conditions. Studies from Emilia-Romagna [57] and Lombardia [43] have found an increase in out-of-hospital mortality due to cardiac causes in the first wave of the pandemic. While Tuscany was relatively less affected by COVID-19 compared to other regions in Italy, complementary studies that examine cardiac conditions care pathway process measures such as door-to-balloon time [58] can better elucidate interpretations. For example, Rossi and colleagues found that in-hospital mortality rates for AMI patients did not change significantly after COVID-19 in Brescia despite an increase in door-to-balloon time, pointing towards a threshold of care management of keeping this measure to 95 min for AMI patients [59]. Third, we were limited to the data available and utilized the percentage of hospital mortality and average length of stay to serve as indicators for quality-of-care management. Future studies may consider collecting data that support measures such as 30-day readmission rates and post-discharge mortality rates to further enhance interpretations of quality-of-care management in hospital.

The COVID-19 pandemic was associated with substantial reductions in hospitalization rates in Tuscany, Italy, but was not associated with changes in hospital length of stay and percentage in-hospital mortality for patients hospitalized for AMI, stroke, and heart failure. The full impact of delayed treatment for these time-sensitive acute conditions remains unclear.