This study was conducted at Peking University Third Hospital, a large tertiary public hospital in Beijing, China. Three CT scanners manufactured by United Imaging Healthcare Co., Ltd. (Shanghai, China) were included: the uCT790, installed in January 2019, primarily used for routine examinations such as chest, spine, and maxillofacial imaging, with an average daily operating time of 18 h; the uCT960+, installed in February 2022, also serving routine examinations including vascular, chest, and pelvic studies, with an average daily operation of 14 h; and the uCT860, installed in November 2023, dedicated to emergency examinations, operating 24 h per day and most frequently used for pelvic, head, and chest imaging. Collectively, these devices represent a spectrum of routine and emergency clinical applications, capturing diverse operational and workload conditions. A failure event refers to the event where equipment parts are replaced.

Maintenance and repair records were retrieved from the hospital’s equipment management database, covering the period from installation to December 2024. Each record of service work order included the following information: work-order type, report time, dispatch time, engineer-arrival time, service-completion time, and failure description. These records were reviewed and cleaned to ensure consistency. Failure events were coded as 1 and non-failure maintenance events as 0. To construct failure time series, data were aggregated at three temporal resolutions—monthly, bi-weekly, and weekly—allowing assessment of model robustness across different time scales.

For the reliability analysis, data from the second full calendar year following each scanners’ installation were examined. In 2020, the uCT790 generated one repair record and one preventive-maintenance record. In 2024, the uCT860 produced ten repair records and two preventive-maintenance records. In 2023, The uCT960+ generated eleven repair records and five preventive-maintenance records. For failure forecasting, data from installation through 2024 were analyzed. Across this period, the uCT790 generated fifteen hardware-related repairs records, the uCT860 produced nine, and the uCT960+ yielded eighteen.

The three CT scanners showed distinct temporal patterns of hardware failures (Figure 1). The uCT790 showed an inverted-V trajectory, rising from one failure in 2019 to a peak of five in 2023, then declining to two in 2024; the most concentrated activity occurred in July 2023, when three failures were recorded. The uCT860 showed a steady upward trend, increasing from three failures in 2023 to six in 2024, with its highest monthly count of three failures in March 2024. The uCT960+ showed relative stabilization, recording four failures in 2022 and plateauing at seven failures annually in 2023 and 2024. Within these years, recurrent monthly peaks of two failures were observed, notably in July 2023 and repeatedly throughout 2024.

Figure 2 illustrated the marked differences in the hardware-failure profiles of the three scanners. For the uCT790, failures were dispersed, with the optical coupler (27%), X-ray tube (20%), and power-related issues (20%) being the most common, alongside contributions from the balance-correcting scanning bed and the temperature control unit. By contrast, the uCT860 showed a highly concentrated failure pattern, with scanning-table faults accounting for more than half of all failures (56%), while other issues—including image artifacts, keypad faults, respiratory-detection module errors, and electrostatic wrist-strap problems—each constitute only 11%. The uCT960 + showed a relatively even distribution of faults, with the scanning bed (28%), respiratory monitoring module (22%), power-related faults (17%) being predominant, while remaining components such as the camera, heat sink, laser light, and leads contribute a certain proportion.

The main reliability and maintenance metrics for the three CT systems are shown in Table 1. The uCT960+ had short cumulative downtime (278.11 h) but the highest number of failures of 11 incidents. The uCT860 accumulated the longest downtime (300.4 h) despite a moderate failure count of 10 incidents. The uCT790 had the both the shortest downtime (205.97 h) and the fewest failures (only one incident). In terms of MTBF, the uCT790 performed best (8554.03 h), while the uCT960+ ranked lowest (771.08 h); however, the differences among the three scanners were not statistically significant (all P-values >0.05). Regarding repair efficiency, the uCT860 had the shortest MTTR (6.595 h) and the lowest MMT (6.686 h). The uCT960+ demonstrated the fastest support response, as indicated by the shortest MLDT (3.471 h).

For the uCT790, observed and predicted failure counts were almost always zero, with only minor deviations in September and December, yielding an MAE of 0.33 (MAE/mean >50%). The uCT860 had low observed failures but highly variable predictions—for example, three predicted failures in July versus zero observed, and two observed failures in October versus zero predicted—resulting in a higher MAE of 1.5 (MAE/mean = 45.4%). The uCT960+ consistently showed observed and predicted counts close to one, achieving an MAE of 0.33 (MAE/mean = 39.9%) and demonstrating the highest predictive accuracy (Table 2).

During October–December 2024, the uCT790 recorded zero failures in all observed and predicted half-month intervals, yielding an MAE of 0 (MAE/mean <10%). The uCT860 showed a one-count over-prediction in the second half of October 2024 (MAE 0.33; MAE/mean >50%), while the first half of November matched exactly (MAE 0; MAE/mean < 10%). The uCT960+ deviated only in the second half of December 2024, where one observed failure was under-predicted (MAE 0.17; MAE/mean >50%); all other intervals matched exactly (Table 3).

Across week 48th to 53rd of 2024, observed and predicted counts remained stable (Table 4). The uCT790 and uCT860 had zero failures throughout, yielding an MAE of 0 and indicating perfect predictive accuracy (MAE/mean <10%). The uCT960+ matched zero for week 48th–52th, with a slight under-prediction of one failure in week 53rd (MAE 0.17; MAE/mean >50%).

Our results were consistent with the findings from previous studies emphasizing the susceptibility of CT systems to hardware degradation under high workloads. For instance, X-ray tubes and slip ring assemblies are frequently considered as primary failure-prone components due to heat accumulation and continuous electrical transmission [14]. Similar to our finding in the uCT790, which despite the highest MTBF still experienced frequent X-ray tube failures, this highlights the importance of complementing system-level reliability metrics with component-level analysis. The uCT790 showed a fluctuating increase in failures from 2019 through 2023, peaking in 2023 before declining in 2024. This trajectory likely reflects the cumulative burden of operating hours, with progressive aging and prolonged use gradually exposing latent weaknesses, while the 2024 decline suggests the benefits of recent maintenance or component replacements. The uCT790 is frequently used for chest imaging, where the X-ray tube must continuously adjust its output power to accommodate variable tissue densities, leading to heat accumulation and accelerated wear.

The uCT860s high rate of scanning-table failures mirrors earlier showed that mechanical subsystems, especially patient couches subjected to frequent motion and more frequently. The uCT860, used exclusively for emergency examinations, showed a steady year-on-year increase, reaching six failures in 2024. This escalation reflects its round-the-clock, high-intensity workload, with scanning-table failures accounting for 56% of the total and highlighting the substantial stress imposed by continuous patient throughput [15]. The predominance of pelvic examinations further contributes to repeated mechanical and electrical strain on the scanning bed, accelerating its wear. In many vascular CT/MRI imaging protocols, performance depends heavily on auxiliary systems such as respiratory monitoring modules and ECG gating. Movement control and peripheral monitoring are among the more failure-prone subsystems in high-use settings. The relatively stable performance of the uCT960+, followed by a rise in failures in later years, reflects progressive wear of auxiliary components such as respiratory monitoring modules, corroborating earlier observations of peripheral system vulnerabilities in vascular imaging protocols [16, 17].

The uCT960+, used mainly for routine rather than emergency imaging, showed relative stability in 2022 and 2023, followed by a sharp rise to seven failures in 2024. This pattern is most plausibly attributable to the combined effects of increasing service duration and progressive component degradation. The uCT960+, often used for vascular detection, also strains components like the respiratory monitoring module. Generally, equipment failure frequency and causes are greatly influenced by usage patterns. To enhance operational efficiency and reliability, customized maintenance and improvement measures should be implemented based on each device’s characteristics and failure patterns.

The reliability analysis showed that the uCT790 had the lowest MTBF (602.9 h) compared with the uCT860 (809.1 h) and the uCT960+ (670.2 h), indicating shorter operating intervals between failures, most plausibly reflecting its longer service duration and cumulative aging effects. By contrast, the higher MTBF of the uCT860 suggested greater inherent stability, likely due to more advanced design, improved manufacturing quality, or implementation of robust preventive maintenance programs [18]. In terms of MTTR, the uCT90 required the longest repair time (19.4 h), exceeding the uCT860 (18.7 h) and the uCT960+ (13.63 h), suggesting that prolonged repair time of the uCT790 may be attributable to hard-to-source components or more complex service procedures. By comparison, despite a higher failure count, the uCT960+ achieved the shortest MTTR, reflecting efficient corrective maintenance, probably facilitated by robust diagnostics or a modular architecture that enabled rapid part replacement [19, 20].

In forecasting analyses, the uCT860 showed a markedly higher monthly MAE (1.5 h) compared with the uCT790 and uCT960+. This poorer performance is largely explained by its short installation history (since late 2023), which limited the training set and precluded detection of long-term failure patterns, as well as the stochastic nature of failures arising from its continuous emergency use. By contrast, the uCT790 and uCT960+ had longer operational histories and more regular workload patterns, allowing the model to capture steadier trends and achieve lower MAE values. As temporal resolution increased from monthly to bi-weekly and then to weekly, the MAE decreased across CT scanners. Finer time steps reduced data volatility and randomness, enabling the model to capture short-term failure patterns with greater precision. For example, the uCT860s MAE declined from 1.5 (monthly) to 0.33 (bi-weekly) and 0.17 (weekly), demonstrating improved predictive accuracy for stochastic events at shorter intervals. Overall, predictive accuracy varied strongly by temporal granularity: weekly forecasts achieved the highest precision, bi-weekly forecasts were intermediate, and monthly forecasts were least accurate. However, weekly predictions were dominated by zeros, raising concerns about zero-inflation bias and increased labor costs for model updating. A balanced assessment of predictive stability and resource expenditure indicated that bi-weekly forecasts provided the most cost-effective approach. Although the MAE/mean observed of the groups is greater than 50%, due to the extremely sparse nature of the data, the relative error ratio naturally increases. At this point, the absolute error MAE is more practically valuable.

This study has several limitations. First, the historical failure records were relatively short for some systems—particularly the uCT860 and uCT960+—which may have constrained the accuracy and stability of failure forecasts. Second, although we examined key reliability and maintainability metrics, the analysis did not incorporate all contextual variables that may influence equipment performance, such as differences in operator practices, variations in maintenance staff expertise, or supply-chain delays for spare parts. Third, the prediction model was applied to a single institution and a single manufacturer’s CT systems, which may limit the generalizability of findings to other vendors, institutions, or health system contexts. Finally, the forecasting models, while useful for short-term trend prediction, are inherently data-driven and do not explicitly account for unobserved clinical or environmental shocks, such as sudden changes in workload during epidemics. The study has strengths. To our knowledge, this is one of the large-scale investigations to integrate reliability metrics, failure patterns, and predictive modeling across multiple CT scanners within a real-world tertiary hospital. The analysis not only identified distinct equipment-level failure trajectories but also linked these to component-level vulnerabilities and usage patterns, offering mechanistic insights beyond simple failure counts. Moreover, by reparameterizing Prophet into a Poisson-Prophet framework, we adapted a widely used forecasting tool to accommodate the sparse and discrete nature of medical device failure data, thereby enhancing methodological rigor [21]. Finally, the combination of system-level reliability metrics with predictive modeling provides actionable evidence to guide proactive maintenance scheduling, resource allocation, and long-term equipment management in high-demand clinical environments.