Predicting medical device failure: a promise to reduce healthcare facilities cost through smart healthcare management

Predictive model framework

This study considers five types of healthcare facilities under the Ministry of Health, Malaysia. They are four types of hospitals: state, major, minor, non-specialist, and one special psychiatric institution, which is equivalent to 15 healthcare facilities. A predictive model was developed based on the medical device data (Engineering Services Division Ministry of Health, 2018) from government hospitals in Perak (west coast of Malaysia peninsula). Perak state hospital is equipped with 990 beds, two major specialist hospitals with 608 and 548 beds, two minor specialist hospitals with 305 and 250 beds, nine non-specialists or district hospitals with 50 to 160 beds, and 1,800 beds for a special psychiatric institution. The state hospital provides 15 specialist services and designated sub-specialists based on the region, and clinical service is managed by clustering in their respective areas. The 15 medical specialist services cover and are not limited to general medicine, general surgery, pediatrics, orthopedics, obstetrics and ophthalmology, ENT (otorhinolaryngology), emergency medicine, psychiatry, dental, dermatology, and nephrology. From these 15 facilities, there were 12,214 medical device units with active and inactive critical medical devices from 1997 to April 2021. The medical device maintenance at these 15 facilities is currently performed by a concession company appointed by the government in a long comprehensive contract. A challenge is encountered in gathering, integrating, maintaining, processing, and analyzing various types of medical data. It is too complicated and inefficient to handle using existing database management systems because it consists of big healthcare data. Although computerized maintenance management can store a vast number of data, difficulties occurred when clinical engineers could not provide myriad maintenance data into a robust tool that enables the implementation of comprehensive maintenance strategies. The utilized medical device database contains both structured and unstructured data. Structured data is the device’s general information, whereas unstructured data includes routine maintenance service records and troubleshooting actions performed by competent personnel (i.e., clinical engineers). Utilizing unstructured data necessitates a laborious pre-processing phase to clean and organize the data.

Predictive model development

Government hospitals in Malaysia used a web-based database to record and store medical device history since 1997. The web-based data for this analysis is asset and services information system (ASIS), which spans from the system’s inception in 1997 to April 2021. Figure 1 depicts the overall proposed framework where data on 44 types of critical medical devices are extracted from 15 healthcare facilities (comprised of 8,294 devices). The pre-processing data stage is executed, where 17 input parameters are selected based on literature review findings. All 17 features from 8,294 devices are extracted to be fed to the proposed predictive model to anticipate the likelihood of first failure of the medical devices.

Normalization is required in the pre-processing stage, where this process will ensure all features are within the same scale and range. After normalization, data is combined with categorical data as an input parameter or features to the predictive model. Normalization returns the vector-wise $z-score$ of the dataset with center zero (0) and a standard deviation of one (1). Normalization operates on each column of data separately, and the below equation is used for $z-score$ of a value $x$ (Zamzam et al., 2021);

$$z-score={\displaystyle \frac{x-\mu }{\sigma }}$$

$$\sigma =\sqrt{{\displaystyle \frac{{\sum }_{i=1}^n{\left(x-\mu \right)}^2}{n}}}$$

(x = value, μ = mean, σ = standard deviation, n = highest probability estimate of the population’s standard deviation)

The data point distance from the mean and standard deviation is the measurement for $z-score$. Returning as a vector and matrix, standard deviation of $x$ and mean of $x$ are used to calculate the $z-score$. After normalization is executed, a normalized value is imported into the software.

The following process flow is to identify the response classes boundary for classes 1, 2, and 3. The class is divided using an arbitrary technique based on the pattern in the data. The model is trained on five algorithms using the cross-validation technique with 80% training and 20% testing data segregation. Observing the confusion matrix, the model performance is examined, and the values for recall, precision, specificity, F1 score, and AUC are calculated. We perform sensitivity analysis to further optimize the developed predictive model in determining the most significant features for medical device failure prediction. The proposed model is then tested using a new test dataset to predict failure classes and the outcomes on cost impact. A new proposed maintenance schedule and replacement plan will be discussed further in the discussion section.

Machine learning and deep learning application

In this stage, ML and DL techniques are explored using five algorithms and three optimizers for both approaches, respectively (Fig. 1). Both ML and DL techniques were compared with the 17 features embedded in the model. A sensitivity analysis is performed to optimize the reliability and rank the significant features in determining the best predictive model. Decision trees, naïve Bayes, support vector machines, ensemble classifiers, and neural network algorithms are used for ML applications. Meanwhile, the stochastic gradient descent with momentum (SGDM), root mean square propagation (RMSProp), and adaptive moment estimation (Adam) are applied for DL. A support vector machine in ML splits data into classes by locating the optimal hyperplane that divides each point into its corresponding category. Conversely, numerous weak learners’ outputs are combined into one accurate ensemble model using ensemble classifiers algorithm by boosting the maximum number of splits and learners. A similar tree model is also applied in decision trees with responses predicted by following the root to leaf node and containing responses in true or false conditions. Gaussian distribution with a mean and standard deviation is used in naïve Bayes to simulate the predictor distribution within each class. In addition, a feedforward fully connected neural network is utilized, which has a connected layer with each fully linked layer that multiplies the input by a bias vector and a weight matrix.

Moreover, a convolutional neural network (CNN) for DL is used with SGDM, RMSProp, and Adam as the training options. SGDM optimizer specifies the momentum value using momentum training options, and RMSProp has a decay rate option using squared gradient decay factor. Adam optimizer is propounded with decay rates of gradient besides squared gradient moving averages using gradient and squared gradient decay factor. As a result, the SGDM optimizer can oscillate along the precipitous descent, leading to the best result. One method to decrease this oscillation is to include a momentum term in the parameter update as specified in the SGDM equation below (Dubey et al., 2020; Essai Ali & Taha, 2021; Setiawan et al., 2019; Sultana et al., 2021):

${\theta }_{\ell +1}={\displaystyle {\theta }_{\ell }-\alpha \mathrm{\nabla }E({\theta }_{\ell })+\gamma \left({\theta }_{\ell }-{\theta }_{\ell }-1\right)}$

(γ = current iteration’s contribution from previous gradient step, $\alpha$ = learning rate, $\ell$ = iteration number, $\theta$ = parameter vector, $\mathrm{\nabla }E(\theta$) = loss function).

The application of learning rates that vary by parameter and may automatically adjust to the optimized loss function shall enhance the network training. RMSProp comes into place where it upholds a moving average of the parameter gradients’ element-wise squares with a decay rate of the moving average identified as β2 and is applied in the below equation (Dubey et al., 2020; Essai Ali & Taha, 2021; Setiawan et al., 2019; Sultana et al., 2021):

$${\nu }_{\ell }={\beta }_2{\nu }_{\ell -1}+\left(1-{\beta }_2\right){\left[\mathrm{\nabla }E\left({\theta }_{\ell }\right)\right]}^2$$

${\theta }_{\ell +1}={\displaystyle {\theta }_{\ell }-{\displaystyle \frac{\alpha \mathrm{\nabla }E\left({\theta }_{\ell }\right)}{\sqrt{{\nu }_{\ell }}+\in }}}$

( ${\beta }_2$= the moving average’s rate of decay, $\in$ = small constant is added to prevent zero division).

Besides, RMSProp has similar characteristics to Adam, provided Adam has the added momentum term. It maintains a moving average, element by element, of the parameter gradients and their squared values using the below equation (Dubey et al., 2020; Essai Ali & Taha, 2021; Setiawan et al., 2019; Sultana et al., 2021):

$$m_{\ell }={\beta }_1{m}_{\ell -1}+\left(1-{\beta }_1\right)\mathrm{\nabla }E\left({\theta }_{\ell }\right)$$

$${\nu }_{\ell }={\beta }_2{v}_{\ell -1}+\left(1-{\beta }_2\right){\left[\mathrm{\nabla }E\left({\theta }_{\ell }\right)\right]}^2$$

$${\theta }_{\ell +1}={\theta }_{\ell }-{\displaystyle \frac{\alpha m_{\ell }}{\sqrt{{\nu }_{\ell }}+\in }}$$

( ${\beta }_1$ = gradient decay rate).

A decay rate value of β1 and β2 can be specified using the gradient decay factor. Adam optimizer uses a moving average, and network parameters are updated using the equation. When gradients over several iterations are comparable, employing a moving average of the gradient allows the parameters to change and gain momentum in a particular direction. The parameter updates also decrease in size if the gradient is primarily noise-based because the moving average of the gradient shrinks. A technique to counteract a bias that occurs at the start of the training is also included in the whole Adam network.

Features selection

Numerous critical medical devices are discussed in the Malaysia Standard of Good Engineering Maintenance Management of Active Medical Devices in MS2058: 2018 (Department of Standards Malaysia, 2018). They are categorized into three main groups: therapeutic, diagnostic, and laboratory. Besides, the devices are grouped into critical equipment and patient support equipment. This article selects the critical medical devices as listed in MS2058: 2018. From this list, 72.4% of therapeutic and 48% of diagnostic devices are categorized as critical devices, whereas these two categories also increased efficacy, complexity, and tendency for adverse effects (Curtis, Tzannes & Rudge, 2011). Despite its criticality, in 2020, the diagnostic imaging equipment segment held the greatest market share, significantly impacting the medical device maintenance market globally (Markets and Markets, 2021). All devices under the laboratory group are categorized as patient support equipment and not classified as critical medical devices. Hence, by eliminating inactive medical devices and laboratory medical devices, the total number of 8,294 active medical devices. All 44 types of critical devices are located at critical locations such as operation theatre (OT), ICU, accident and emergency (A & E), and wards. A total of 34.68% of the devices are infusion pumps, and 14.93% are physiologic monitoring systems which are located in various areas in Perak hospitals. Other devices below 10% in percentage are radiographic systems, drills bone, cystoscopes, colonoscopes, colposcopes, laparoscopes, mobile radiographic/fluoroscopic, dental radiographic, surgical hand drills, injectors, lithotripters, pacemakers, peritoneal dialysis units, resuscitators, stimulators, and ultrasonic. High-end medical devices or other categories are small in numbers and primarily located in X-ray Department. These include Radiographic/Fluoroscopic Systems Angiographic/Interventional, Radiographic/Fluoroscopic Systems General-Purpose Radiographic Units Mammographic, Scanning Systems Computed Tomography, Scanning Systems Magnetic Resonance Imaging, Full-Body. This high-end equipment is located at State Hospitals, Major Specialist Hospitals, and Minor Specialist Hospitals with medical imaging practitioners or Radiologist specialists.

We have selected 17 input parameters or features to be fed and tested in ML and DL predictive frameworks. Details on each parameter are summarised in Table 1, where some features are entered numerically, and others are provided to the ML and DL framework in a categorical manner.

Failure classes

The classification problem in supervised ML is overcome by defining response classes before execution. Hence, based on the tabulated data extracted from ASIS, an arbitrary technique is used to create three classes based on the available dataset. Then, the best balance classes are selected, and the response classes for critical medical devices are subdivided into classes 1, 2, and 3. With the capability of supervised machine learning analyzing retrospective data, classifying critical medical devices is beneficial for more effective PPM planning, management replacement plan, and strategic annual budget preparation. As for newly purchased critical medical devices, the proposed model can evaluate its performance throughout its lifespan based on the brand, model, failure history, and other features used during the training stage. Five algorithms are utilized in ML: decision trees, naïve Bayes, support vector machines, ensemble classifiers, and neural networks. In addition, a cross-validation technique with segregation of 80% training and 20% testing is used.

The proposed predictive model uses 17 features, as tabulated in Table 1, to distinguish the devices into three classes. Class 1 is defined as unlikely to fail within the first 3 years from the purchase date, while class 2 is for devices that are likely to fail within 3 years. Class 3 is for devices likely to fail more than 3 years after purchase, as tabulated in Table 2. Medical devices in class 1 are less critical and have low complexity in maintenance than in classes 2 and 3. Class 1 can be identified as the least problematic devices, including newly purchased ones. Class 2 is a matter of concern since the failures are detected closer to the purchase date. The class 3 devices are classified as prone to failure after 36 months of purchase date. The scheduled maintenance frequency is conducted as per the manufacturer’s recommendation in regular maintenance practice. There is no consideration has been made based on the failure history data. Therefore, with the proposed predictive framework in this study, new recommendations shall be made where maintenance frequency and cost can be reduced based on the forecasted analysis attained from the proposed model. Out of the total 8,294 critical devices in 15 different hospital categories, 49.51% of devices were classified as class 2, while the remaining 26.89% and 23.58% are categorized as Class 1 and 3, respectively. Thus, the yearly budget allocation for class 1 can be reduced and should be reallocated to classes 2 and 3.

Parameter tuning and optimization

Parameter tuning and optimization are executed to improve classification performance, accuracy and introduce a unique identity to the model. Ensemble classifier outperforms other algorithms with an accuracy of 77.90%, followed by decision trees with 76.30% after parameter optimization, as explained in Table 3. Receiver operating characteristics (ROC) curves and AUC were measured where the higher the AUC values (as it is closer to ‘1’) indicate an excellent predictive model (Shiferaw, Bewket & Eckert, 2019). In this study, Ensemble classifiers attained an AUC of 0.89, decision trees 0.85, while both support vector machines and neural networks attained 0.82. Meanwhile, naïve Bayes only achieves 0.80 for AUC values. Parameter optimization increased the model accuracy for four algorithms. However, parameter optimization for decision trees remains at an accuracy of 76.30% before and after optimization. Hence, ensemble classifier performs best with the highest recall, precision, specificity, and F1 Score values after parameter tuning and optimization. As compared to the DL technique, SGDM optimizer denotes the highest accuracy compared to RMSProp and Adam optimizer. There is a reduction in performance accuracy from 77.90% for ML to 70.33% for the DL model. However, DL has the advantage of shorter training time than ML. DL requires 1 min 5 s to complete the training progress, much faster than 11.49 min using ML.

A model optimization is performed for all algorithms and optimizers in both techniques. A kernel scale is optimized in support vector machines algorithm from automatic to one, producing 65.90% instead of 65.80% during pre-optimization, as in Fig. 2. The decision tree algorithm’s accuracy remains at 76.30% when a maximum number of splits is tuned from 100 to 1. A different number of splits gives zero impact to the model but demands a higher training time from 4.1609 to 33.053 s. As for naïve Bayes, the accuracy increased from 57.50% to 59.80% by changing the categorical predictors to Gaussian parameters with a training time increase from 40.397 to 366.74 s. Meanwhile, the ensemble classifier improves its accuracy from 76.00% to 77.90% during the optimization stage when the maximum number of splits is reduced from 6,635 to 20, and ensemble method is set from Bag to Ada Boost. The number of learners is retained at 30 after optimization, and learning rate is set to 0.1 with an increasing training time from 16.575 to 689.92 s. Besides, a number of fully connected layers is optimized from 1 to 3, and first layer size is reduced from 100 to 10 for the neural network algorithm, improving the accuracy from 69.60% to 69.90%. However, neural network algorithm in supervised ML denotes the longest training time of 57,685 s.

The model optimization is applied for DL, where a tuning of mini-batch size from 128 to 100 with an increase of epoch from 30 to 60 gives a different performance impact to the model. SGDM optimizer uses a mini-batch size of 100, and a maximum epoch of 60 attained an accuracy of 70.33%, slightly lower than the ensemble classifier. The diminution of mini-batch size from 128 to 100 increases the model accuracy from 65.26% to 68.03% for RMSProp and 68.03% to 70.33% for SGDM optimizer, as in Fig. 2. Nevertheless, the optimization process reduced the accuracy from 68.88% to 68.76% for Adam optimizer. A training progress graph interprets a DL technique, representing the accuracy of each unique mini-batch. The model performance can be monitored in real-time and model progress can be stopped at any time. Figure 3 shows RMSProp takes the longest elapsed time of 1 min 13 s, followed by 1 min 12 s for Adam and 1 min 5 s for SGDM optimizer, respectively. The model accuracy is plotted in “blue” while model losses are plotted in “orange.” Initially, when the epoch increases, the losses are reduced and continually saturated until it reaches the maximum epoch of 60. The training progress is stopped at epoch of 60 since the progress is saturated and gives an insignificant impact if the process continues.

Predictive model performance and evaluation

The quality of the characteristics given to the algorithm determines the accuracy of the model’s predictions. The overall accuracy and Kappa coefficient are two extensively used indicators for thematic accuracy controls on error matrix (Garcia-Balboa et al., 2018). In addition, precision and recall are two relevant metrics used for evaluating prediction accuracy (Teo et al., 2020). The recall performance measures the trustworthiness of the result or ability to classify positive outcomes at a true positive rate. At the same time, the precision demonstrates the predictive or positive values which correctly predicted (Kareen, 2020). Both precision and recall are frequently at odds, and boosting precision reduces the recall with text retrieval concepts representing both calculations (Ghorbani et al., 2020). A recall is a number of relevant features in the selected subset divided by the total number of relevant features. For precision, it is divided by the total number of features specified in the dataset (Santra & Christy, 2012).

Analyzing and comparing models based on recall and precision is time-consuming; thus, employing the F1 score method is another viable result (Ghorbani et al., 2020). The F1 score is a harmonic average of precision and recall, consider both metrics and evaluate the model’s accuracy and dependability. The confusion matrix is the most common method of reporting on the thematic accuracy of geographic data (Garcia-Balboa et al., 2018; Santra & Christy, 2012). The true class is represented by the rows of the confusion matrix, while the columns represent the predicted class in the 3 × 3 matrix. Correctly classified observations are expressed by diagonal cells; in contrast, erroneously classed observations are represented by off-diagonal cells with TN as a true positive, FP as a false positive, FN as a false negative, and FP as a false positive (Kareen, 2020; Zamzam et al., 2021). An equation or an evaluation metric by Hameed et al. (2021) is used to calculate accuracy, recall, precision, specificity, and F1 score based on a confusion matrix (Kareen, 2020; Teo et al., 2020). The below equation is calculated based on confusion matrix values and is summarized in Table 3. Ensemble classifier has the best performance with 74.67%, 75.39%, 87.60%, and 75.03% for recall, precision, specificity, and F1 score values, as explained in Table 3 previously.

$$Accuracy={\displaystyle \frac{Totalnumberofcorrectlyclassified}{Totalnumberofobservation}={\displaystyle \frac{TP+TN}{TP+FP+TN+FN}}}$$

$$Sensitivity/Recall={\displaystyle \frac{TP}{TP+FN}}$$

$$Precision={\displaystyle \frac{TP}{TP+FP}}$$

$$Specificity={\displaystyle \frac{TN}{TN+FP}}$$

$$F1Score={\displaystyle 2\times \frac{Precision\times Recall}{Precision+Recall}}$$

Sensitivity analysis

Table 4 describes the sensitivity analysis and evaluates the features’ effect on the model performance. Leave one out technique is performed by calculating misclassification and then dividing with all feature errors to obtain the ranking ratio (Chen et al., 2020; Gazzaz et al., 2012; Pastor-Bárcenas et al., 2005). The ratio is ranked in descending order to highlight the impact of the features on the model. Next, five sensitivity analysis techniques, namely Leave One Out, MRMR, Chi2, ANOVA, and Kruskal Wallis, are applied to compare the ranking for all features. MRMR utilizes the minimum redundancy maximum relevance algorithm to rank the characteristics in order. A chi-squared algorithm uses a chi-square test, an ANOVA uses one-way variance analysis, and Kruskal Wallis performs a hypothesis with the same median from the population (The MathWorks Inc, 1994–2021, 2022). These three algorithms ranked the features by −log(p) scores, where the p-values denote the chi-square test statistics. The five highest-ranking features for every technique are underlined in Table 4, and the features are analyzed in the software to achieve better accuracy. Of the 17 features, only eight are identified as the most significant to obtain the highest accuracy. If more than eight features are selected, the accuracy decreases.

Figure 4 compares sensitivity analysis techniques when the eight most significant features are selected to develop the model. The features are age, service support, asset condition, maintenance complexity, total downtime, maintenance cost, model, and purchase date. The graph demonstrates ANOVA and MRMR outperform other techniques with 79.20% accuracy, followed by 79.00% for Chi2. Kruskal Wallis and the Leave one out technique attained 78.10% and 77.50% accuracy, respectively. The model is improved from 77.90% using 17 features to 79.20% accuracy with eight features after sensitivity analysis.

Proposed predictive model using machine learning

The ML predictive model is improved from 77.90% to 79.20% accuracy after sensitivity analysis, as described in the result section. The ML requires more training time of 11.49 min compared to only 1 min 5 s using DL when all 17 features are embedded in the model development. After eight significant features are imported with a learning rate of 0.01 and iterations of 50, the ML model improves its performance to 79.50% accuracy, 76.05% recall, 77.43% precision, 88.36% specificity, and 76.73% for F1 Score indicator. The tuning of learning rate and iterations enhanced the model, with training time reduced from 11.49 to 7.908 min after sensitivity analysis. Although ML has better accuracy than DL, it requires more training time. The DL technique predicts more extensive data better, is primarily used in time series or image data and requires less time than ML. The result concludes that ML performs better in accuracy and all other performance indicators than DL. This best-optimized ensemble classifier model uses Ada Boost with a maximum number of splits of 20 and a learning rate of 0.01 to yield 79.50% accuracy, as shown in Table 5. The proposed model is expected to improve the current system toward smart healthcare management.

The model predicts classes 1, 2, and 3, which represent the time to the first failure event for an effective maintenance schedule and to utilize budget allocation. As compared to the other related works on medical device performance prediction, Kovačević et al. (2020) in infant incubators study predicted the device functionality and classified two different classes: accurate and faulty class with an accuracy of 98.5%. A similar methodology is applied by Badnjevic et al. (2017) for a mechanical ventilator. Using performance parameter values, a defibrillator study achieved 100% accuracy in the Random Forest classifier to predict positive: for devices that passed inspection or negative: for faulty devices (Badnjević et al., 2019). Meanwhile, Hrvat et al. (2020) attained 98.06% accuracy based on a conformity assessment where the outcomes are identified as pass or fail for infusion and syringe pumps. This article has significantly contributed to the medical device reliability assessment research by including 44 types of critical medical devices during model development and analyzing significant cost reduction after implementing the predictive model. To the best of our knowledge, no previous studies proposed a predictive model for anticipating the likelihood of a device’s first failure after the purchase with cost analysis and comparison between ML and DL techniques.

Characteristics for classes, schedule maintenance, and replacement plan

Classes 1, 2, and 3 have different properties, and the boundaries are set based on the patterns from the vast data. The present service contract fee per month in Malaysia is calculated by multiplying the device purchase cost and rate of fee, with the rate of fee defined in percentage based on the type of device. High-end equipment has a higher fee rate equivalent to 19.25%, and the lowest rate is 4.95% per year for small devices (Engineering Services Division Ministry of Health, 2018). The fee includes PPM and corrective maintenance (CM) performed by Concessionaire. It is calculated in lump sum fees regardless of the number of failures events the devices encountered throughout their lifespan. A flat rate is imposed from the purchase date to the end of the device life. As a result, the Concessionaire gains a higher profit during the early age of devices, and the profit margin is reduced and approaches breakeven as the age rises.

$$Servicecontractfee/month={\displaystyle \frac{Purchasecost\left(MYR\right)\times Rateoffee\left(\mathrm{\%}\right)}{12}}$$

Poor maintenance, planning, and management are the most common causes of medical device failures (Arab-Zozani et al., 2021). Countries have different approaches to organizing maintenance and operational budget planning for medical devices. Saudi Arabia uses life cycle cost estimation in making decisions (World Health Organization, 2011a), and Turkey divided medical devices into technological groups to calculate cost distinctly (Bektemur et al., 2018). In the UK, they use purchasing, donations, replacement, and disposal policies to decide where, what, and when to procure medical devices. The replacement budget per year is implied by dividing the device’s current prices by lifetime (Temple-Bird et al., 2005). The New Delhi Medical Equipment Maintenance policy uses the maintenance cost index as an indicator by dividing maintenance cost by capital cost. The maintenance cost values should not increase by 80% of the capital cost of equipment (Kumar, 2012). The United States of America applies the cost of service ratio by dividing the total annual cost for maintenance by the initial cost value. It is used as guidance to improve performance (World Health Organization, 2011b). Overall, at the time of speaking, none of these countries utilize AI applications for maintenance budget planning.

A rule of thumb with 80/20 rule is used for CM and PPM in Stenström et al. (2015), and their results describe PPM represents 10% to 30% of total budget allocation compared to CM. This article’s cost-saving calculation uses 80/20 for CM/PPM yearly and 70/30 for CM/PPM bi-annually for cost estimation analysis. class 1 age ranges between less than 1 year to 30 years, which consists of 2,231 devices in service. There is zero failure, and zero parts cost recorded throughout its lifespan. Throughout the years, PPM is scheduled annually and bi-annually for class 1, with a percentage number of devices are 90.45% and 9.55%, respectively. Therefore, due to its low criticality, PPM frequency is suggested to be reduced from bi-annually to annually with an estimation of cost-saving equal to MYR 199,256.45 per year, as demonstrated in Table 6. On this note, this study proposes a new maintenance strategy, which includes recommendations as illustrated in Table 7. Among others, this study suggests that by reducing the fee rate for class 1 devices, the operational cost can be further reduced. The PPM can be replaced with routine inspection for small devices such as aspirators where the maintenance tasks are minimal; hence the yearly maintenance budget allocation for this group can be reduced for the following year. Maintenance can be executed in-house where only average maintenance complexity with EST is required. In addition, a minimal budget per year is necessary for a loaner or rental costs during downtime for this group of devices.

The biggest class group is Class 2, where 4,107 devices are grouped between 1 to 27 years of age. A total of 83.56% of PPM is scheduled annually, and the remaining are on a bi-annual basis, including 44 types of critical medical devices. Based on the finding of this study, the PPM schedule in class 2 is suggested to remain due to its criticality and risk of failure identified within 3 years of purchase. The devices are expected to fail at any time after being purchased; hence contingency plans or rental devices shall be planned for better service delivery to patients. Higher priority in budget allocation for maintenance and replacement is recommended for class 2 compared to class 1 and class 3. An existing warranty provided by the manufacturer or authorized party is typically within 1 or 2 years after purchase. Therefore, remedial action, such as improving the warranty service to 3 years after purchase, shall accommodate this group’s critical needs. Besides, the cost saving can be maximized by purchasing a 3-year warranty, including breakdown. Hence, zero cost is required under the service contract for the first 3 years, with all PPM and breakdown costs embedded under 3 years warranty. The next group is class 3, with devices likely to fail more than 3 years after purchase, with a lesser risk of failure and complexity than class 2. A percentage of 89.52% of the PPM schedule is planned annually, and 10.48% for bi-annually. PPM frequency is suggested to be reduced bi-annually to annually with an estimated cost-saving equal to MYR 127,074.43 per year, as shown in Table 6. Like class 2, zero cost is required under the service contract for the first 3 years for class 3 devices if PPM cost is embedded under 3 years warranty. The advantage of this alternative is that the service contract fee will only start in 4^th year after class 2 and class 3 devices are purchased. This framework will lead to more cost savings by reducing service contract values.

In addition to the cost-saving demonstrated in Table 6, a breakeven analysis graph can be observed in Fig. 5. This graph shows a case study analysis for a mammographic unit with 21 years of service in the class 3 category. The analysis from the graph depicts that the proposed model has a lesser service contract cost per year compared to the current maintenance practice. With the purchase cost of MYR 422,026.00 allocated by the Government for mammographic equipment, MYR 48,744.00 per year is spent on maintenance services under the present contract. However, if the proposed model is implemented, MYR 41,432.20 per year will be used, which is a 15% savings from the actual cost spent under the present contract. The service contract cost is equivalent to the purchased cost in the 11^th year for present practice and the 12^th year for the proposed model. Therefore, the Government is suggested to replace the equipment no later than 12^th year for better profit management. At the 12^th year, a new device shall be purchased, and unnecessary service contract costs shall be eliminated from 12^th to the 21^st year of service as per the current implementation.

Another strategy to consider is the replacement plan for faulty devices. This study proposes class 2 devices be prioritized for a replacement plan, with 10% of this group’s devices being more than 20 years in service, as tabulated in Table 7. A total of 57.19% of class 2 devices have more than 10 years in service, with a high number of failures throughout their lifespan. Due to the aging factor and a significant number of failure events, replacement with new units should be considered to reduce maintenance costs as the age arises. Class 3 has a similar scenario with 65.18% of devices with more than 10 years in service and should be considered for a replacement right after class 2.

Research contributions

The development of IoT, cloud computing, and AI are continually evolving toward smart healthcare and smart city. The proposed predictive model using AI for medical devices failure prediction is categorized into classes: class 1, 2, and 3. The model accuracy is evaluated by examining other elements in performance evaluation, such as recall, precision, specificity, F1 score, and AUC. The research gap is improved by discovering new scientific findings;

• The proposed model includes 44 types of critical medical devices used in five hospital categories, with 15 healthcare facilities involved, including all critical devices used in the clinical area for patients. The data applied during the training stage includes active medical devices with data on maintenance cost, devices with more than 20 years in service, and sample size up to 8,294 devices resulting in the highest accuracy of 79.50%. Only eight out of 17 features are significant after sensitivity analysis, with a reduction of training time from 11.49 to 7.908 min. A robust model to predict failure in three classes using AI is introduced with fewer features, shorter training time, and comprehensive cost analysis.

• This proposed model can forecast the first occurrence of failure in classes 1, 2, and 3 after medical device is purchased for comprehensive maintenance planning and budget utilization which is currently not in the field of study. The maintenance shall be arranged before the first failure event. Class 1 is the less critical device with zero failure. Class 2 is identified as the most crucial and should be attended to, with the first failure occurrence likely to be within 3 years after purchase. Class 3 is at medium risk; devices are likely to fail after 3 years. To reduce the likelihood of future failures, device replacement for class 2 is in higher priority, followed by class 3. A replacement is proposed by a number of failures and year of service category. This is crucial for any country during crisis management, such as COVID-19 pandemic, where excellent and reliable equipment is highly utilized.

• This article compares the country’s role in organizing budgets while reducing costs in maintenance management. A new PPM schedule and replacement plan frequency are proposed and strategized based on actual needs. Comprehensive strategic management by criticality and devices maintenance history using AI improves maintenance and operational cost. Replacement strategy shall be executed to class 2, then to class 3, depending on the age of the devices and the likelihood of failures.

• Two different techniques between ML and DL were compared, and ML has better performance in accuracy, recall, precision, specificity, and F1 Score. DL has the advantage of shorter training time; however, the accuracy is lower than the ML technique. The accuracy shall be further improved in future work by introducing unstructured maintenance notes written by technical personnel after rectification work is completed.