Discussion

The longitudinal design of the study provides important evidence on the incidence of HAIs in a relatively large sample of Italian LTCFs. The estimated overall HAI incidence of 0.6 per 1,000 resident days complements incidence rates from earlier European studies, underscoring the need for targeted surveillance and prevention strategies. ^{Reference Engelhart, Hanses-Derendorf, Exner and Kramer5,Reference König, Medwed and Pux6}

Beyond its epidemiological contribution, this study extends the application of AI, particularly ML methods, from their predominant use in acute-care hospital settings, ^{Reference Wang, Wang, Wang and Wang13–Reference El Arab17} to LTCFs, suggesting their potential to predict HAIs and inform antimicrobial stewardship in this context as well.

Rather than relying solely on frailty, this method identified a highly stable two-cluster stratification, that captured multidimensional clinical profiles reflecting underlying health status and differential susceptibility to specific types of HAIs. The observed high level of cluster reproducibility indicates that the ML-derived resident stratification was robust and unlikely to result from sampling variability. Such stability reinforces the value of an ML-based approach for reliably defining HAI susceptibility profiles in LTCFs and supports its potential use as an early-warning component for IPC. In addition, the identification of high-risk clusters may enhance antimicrobial stewardship by prioritizing targeted screening and guiding differentiated antibiotic protocols, thereby reducing inappropriate antimicrobial use.

Compared to Group 1, Group 2 included more clinically vulnerable residents with higher exposure to RTIs, UTIs, and skin and soft tissues infections. These infections were significantly more frequent among residents with a high prevalence of underlying health conditions such as disorientation, incontinence, low mobility, hemiplegia, kidney disorder, dementia, malignancies, diabetes, and chronic lung disease. Most of these conditions are well-recognized risk factors for bacterial HAIs, particularly UTIs and skin or soft-tissue infections. The observed co-occurrence of incontinence and UTIs, and of chronic lung disease and RTIs aligns with previous reports and supports the biological plausibility of these findings. ^{Reference Bennett, Tanamas and James3,4,Reference Furmenti, Rossello and Bianco23,Reference Baranowska-Tateno, Micek, Gniadek, Wójkowska-Mach and Różańska24} In contrast, Group 1 residents, characterized by better mobility, greater functional autonomy, preserved cognitive status, and by a lower overall comorbidity burden, showed a higher proportion of COVID-19 infections. Greater mobility within the facility and increased interpersonal contact may have facilitated exposure to respiratory viral pathogens, consistent with patterns described in LTCF COVID-19 outbreaks. ^{Reference McAndrew, Sacks-Davis, Abeysuriya, Delport, West, Parta, Majumdar, Hellard and Scott25} Notably, G1 included residents with clinically significant systemic diseases (as defined in Table 2), underscoring that serious inflammatory conditions may occur despite a relatively low overall comorbidity burden, reflecting substantial heterogeneity across individuals and among different rheumatic disorders. ^{Reference Radner26}

Prior studies have demonstrated that variation in resident frailty, disability, and clinical case mix substantially influences infection risk, prompting efforts to operationalize these differences using administrative measures such as the Case-Mix Index (CMI) and Resource Utilization Groups. ^{Reference Fries, Schneider and Foley27–Reference Mylotte30} However, no consistent association was observed between CMI and infection rates, with the notable exception of a study conducted in the same Italian regions included in the present analysis. ^{Reference Marchi, Grilli, Mongardi, Bedosti, Nobilio and Moro31} Taken together, these findings suggest that traditional administrative measures may not fully capture the multidimensional clinical vulnerability underlying susceptibility to HAIs in LTCFs. Likewise, frailty indices and comorbidity scores summarize vulnerability primarily as an overall burden of deficits, limiting their ability to differentiate distinct clinical profiles within already frail LTCF populations. ^{Reference Clegg, Young, Iliffe, Rikkert and Rockwood32,Reference Charlson, Pompei, Ales and MacKenzie33} In contrast, the unsupervised clustering approach adopted in this study offers a methodologically distinct and complementary perspective: rather than ranking residents along a single continuum of frailty or comorbidity, it integrates functional, cognitive, clinical, and care-related variables simultaneously, without imposing predefined weights or linear assumptions. This data-driven framework may facilitate the identification of latent clinical profiles defined by specific combinations of characteristics, thereby helping to capture heterogeneity that could be masked by unidimensional summary measures.

Consistent with this perspective, the two clusters identified in this study share several underlying health conditions, reflecting the generally high baseline vulnerability of the LTCF population. However, they differ in functional, cognitive, continence, and care-dependency patterns that may better capture infection susceptibility than overall frailty.

This approach aligns with contemporary models of geriatric care, which move beyond the traditional view of LTCF populations as uniformly frail, and instead adopt a more dynamic, risk-adjusted perspective. ^{Reference Fazio, Pace, Flinner and Kallmyer11,Reference Clegg, Young, Iliffe, Rikkert and Rockwood32} In this context, the integration of ML techniques into epidemiological surveillance enhances early warning capabilities and supports evidence-based clinical decision-making through resident-level data stratification. The well-documented severe consequences of HAIs among LTCF residents reinforce the urgency of identifying high-risk resident groups and tailoring preventive interventions accordingly. ^{Reference Koch, Eriksen, Elstrøm, Aavitsland and Harthug34} Additionally, high rates of colonization with multidrug-resistant organisms, as observed in Italian LTCFs, ^{Reference Giufrè, Ricchizzi and Accogli35} further underscore the critical need for proactive surveillance and stratification strategies to mitigate infection risk.

Some limitations of this study should be considered when interpreting the findings. Incomplete resident enrollment due to informed consent requirements prevented a full census in some facilities, potentially limiting representativeness. Such barriers to comprehensive surveillance are well recognized in LTCFs, where resident turnover and consent procedures often hinder systematic data collection. ^{4,Reference Juthani-Mehta and Quagliarello36} Generalisability may be further constrained by the heterogeneity of participating LTCFs and their differing case-mix profiles, as well as by the use of a convenience sample, as previously noted in the H4LS study. ^{Reference Kohler and McGeer37} Reliance on manual data collection rather than routinely collected electronic data in LTCFs constrains the scalability of AI-based approaches. This highlights the need for shared data repositories that integrate resident-level information across facilities. Access to such data would enable the application of information extraction methods such as clustering, to classify facilities based on the proportion of residents at high risk for HAIs. Finally, although formal statistical testing of associations between comorbidities and HAIs was not feasible with the present data set, and the observed cluster-specific infection profiles should therefore be considered as hypothesis-generating rather than confirmatory, these patterns suggest that preventive strategies may benefit from being tailored to residents’ functional and clinical profiles.

In conclusion, this study indicates that AI-based clustering is feasible in LTCFs and may complement traditional infection risk assessment methods. In a field where AI applications remain limited, these findings support further investigation of machine-learning approaches using larger and more comprehensive data sets. By capturing infection-specific risk profiles that extend beyond global frailty, ML-based clustering highlights subgroups with higher HAI incidence and may inform targeted prevention and more efficient resource allocation. Future research should assess the reproducibility of these results across diverse LTCF settings and evaluate the impact of AI-driven risk stratification on resident outcomes, healthcare costs, and infection prevention strategies.