Result of literature search
In total, 165, 52, 153, and 198 studies were obtained from CNKI, Wanfang Database, VIP, and CBM, respectively; 463, 2744, and 154 studies were obtained from PubMed, Embase, and The Cochrane Library, respectively; and 28 studies were obtained through the HTA professional websites and the references of the included studies. All studies were screened strictly in accordance with the inclusion and exclusion criteria. Finally, 40 studies [7-46] were included (2 HTAs, 26 controlled studies, 11 non-controlled studies, and 1 economic study), among which 25 studies were in foreign languages and 15 studies were in Chinese. The literature inclusion process and results are shown in Figure 1.
Results of quality assessment
The quality assessment of HTAs showed the quantity of evidence was small and the quality was low. The quality assessment of the economic study showed the the quality score (89 points) was relatively high. The description of methodology in the qualitative study was not clear and it was difficult to assess the quality. In addition, a survey report didn’t have an appropriate scale for quality assessment. Details of quality assessments for other studies can be found in Appendix 2.
Results of bibliometric analysis (figure 2)
Since publication of the first study, the number of studies published in this field has increased each year. After more than 10 years of development, the application of intravenous compounding robots in China and abroad has been gradually increasing, and the number of relevant studies has reached its peak in the past three years.
Results of comprehensive analysis
Two HTAs [11,12] were included (Table 1); both were published by the CADTH in 2013 and 2016. The results of these two HTAs showed that automation for the preparation of intravenous admixture medicines had certain security and economy. However, because the number and quality of the included original studies were low, it is necessary to update the currently available HTAs or re-evaluate the available data. International (INAHTA, HTAi, ICES, ISPOR), American (AHRQ), European (EUnetHTA), British (NIHR HTA Programme), Canadian (CADTH), Swiss (SBU), and Australian (AGDHHTA) HTA databases do not yet contain any evaluations of intravenous compounding robots.
Thirty-eight original studies [7-10,13-46] (Table 2) were included and published from 2006 to 2020 by researchers from 11 countries, including China (16 studies, 42%), America (6 studies, 16%), Germany (4 studies, 11%), France (2 studies, 5%), Holland (2 studies, 5%), Italy (2 studies, 5%), and Japan (2 studies, 5%). As classified by study type, these 38 studies comprised 24 controlled studies (6 RCTs, 13 contemporaneous NRCTs, and 6 historical NRCTs), 12 non-controlled studies (11 single-arm studies, 1 investigation report), 1 qualitative study (in-depth interview), and 1 economic study. As classified by intervention type, the 38 studies comprised 26 studies that applied intravenous compounding robots and 12 studies that applied ADE (such as automated compounding devices and bidirectional precise infusion solution dispensers). Different models of automated (or robotic) equipment were applied, including the CytoCare System (Health Robotics), APOTECAchemo System (APOTECA), i.v.STATION System (Aesynt), RIVA System (Intelligent Hospital Systems), and IntelliFill i.v. System (Baxter Healthcare) as well as robots independently developed in China, such as WEINAS series robots and Angel compounding robots. The sample size of each study ranged from 10 to 11,865, and the observation time ranged from 2 days to 3 years. The main types of intravenous admixture medicines were anti-tumor drugs (14 studies, 36%), unclassified medicines (14 studies, 36%), antibiotics (2 studies, 5%), and total nutrient admixture (2 studies, 5%).
Effectiveness: Thirty studies evaluated the effectiveness (Table 3), including 21 controlled studies and 8 non-controlled studies. The evaluation indicators were production efficiency, medicine residues, preparation accuracy, preparation errors, preparation failures, error recognition, and solution properties.
Twenty-four studies [7,9,10,13-17,19-23,26,29-31,33-34,38,40-42,46] evaluated the production efficiency, including 18 controlled studies, 5 non-controlled studies, and 1 qualitative study. The data could not be merged because of the different medicine types in each study. However, 12 control studies [13-17,19-22,29-31] showed that the number of preparations in the robot groups per unit time (minimum: 42.13 ± 6.83 bags/hour, maximum: 275 ± 8.52 bags/hour) was significantly higher than that in the manual group (minimum: 26.22 ± 7.52 bags/hour, maximum: 96.6 ± 10.0 bags/hour). The preparation speed of the robot group (minimum: 0.38 ± 0.03 min/bag, maximum: 14.98 ± 0.24 min/bag) was significantly higher than that of the manual group (minimum: 0.55 ± 0.13 min/bag, maximum: 20.21 ± 0.68 min/bag). One non-controlled study  also showed that the robot had good production efficiency, and one in-depth interview  revealed that the robot improved the production efficiency by three to five times compared with the traditional manual preparation. Four studies showed that with increases in pharmacists’ familiarity with robots [10,46] and increases in the quantity of preparations [7,38,46], the production efficiency of the robot groups was further improved. In one controlled study , the number of preparations by the robot reached 70.4% of the total number of preparations in the inpatient ward. Another controlled study  showed that an automated compounding device shortened the order-to-administration time for antibiotic first doses (from 4.5 ± 4.1 to 2.9 ± 2.5 hours, p = 0.009), which was of great significance for the improvement of therapeutic effects and clinical outcomes. However, one controlled study  showed no significant difference in the preparation time before and after introducing the robot, and another controlled study  showed that the average preparation time in the robot group increased by 47% compared with that in the robot group but that the pharmacists’ working time decreased by 76%. Yet another controlled study  showed that the average number of preparations by the robot in 7 hours was comparable with that by the trained and experienced pharmacy staff in 2 to 3 hours; thus, the robot produced limited improvement in production efficiency in practice. Another non-controlled study  showed that although robots can significantly improve production efficiency, the manual pre-processing and post-processing steps were time-consuming and had to be reorganized.
Nine studies [13-15,17,19,20,23,40,46] evaluated medicine residues, including 7 controlled studies, 1 non-controlled study, and 1 qualitative study. Six controlled studies [13-15,17,19,23] conducted a quantitative analysis of medicine residues. The data could not be merged because of the different medicine types in each study, but the results of all studies showed that the rates of residues in the robot groups (0.00%–4.50%) were significantly lower than those in the manual groups (3.67%–50.00%). One non-controlled study  also showed that the use of robots reduced medicine residues, and another in-depth interview  showed that the robots reduced the residual amount of some insoluble medicines and alerted the technicians through an alarm when the residual amount was large. Only one controlled study  showed that the amount of medicine residues in the robot group (compound coenzyme: 0.11 ± 0.01, sodium carbamate: 0.12 ± 0.01) was slightly higher than that in the manual group (compound coenzyme: 0.09 ± 0.02, sodium carbamate: 0.08 ± 0.02); however, the amount in both groups was lower than the hospital inner quality standard.
Ten studies [28,9,23,7,30,44,41,38,35,46] evaluated the preparation accuracy, including five controlled studies, four non-controlled studies, and one qualitative study. Seven controlled studies [28,23,9,7,30,44,41] conducted a quantitative analysis on the preparation accuracy (deviation of less than ±5%). The data could not be merged because of the different medicine types in each study, but the results of six studies [28,9,7,30,44,41] showed that the rates of preparation accuracy in the robot groups (0%–23%) were significantly higher than those in the manual groups (5%–53%). One controlled study  showed that robotic preparations were more accurate and precise (mean absolute dose error of 0.83 for fluorouracil and 0.52 for cyclophosphamide) than manual preparations (1.20 and 1.70, respectively). In an in-depth interview , the vast majority of respondents indicated that compared with manual preparation, the robot improved the accuracy by adjusting the dose by itself. Two non-controlled studies [38,35] showed that the preparation accuracy was related to the viscosity of the liquid and the minimum volume of the dose. The minimum volume of the non-viscous solution and viscous solution that could be accurately prepared (deviation of less than ±10%) was 2 ml and 5 ml, respectively .
Eleven studies [13-15,17,19,26,31,32,37,40,46] evaluated the incidence of preparation errors, including eight controlled studies, two non-controlled studies, and one qualitative study. Seven controlled studies [13-15,17,19,26,31] conducted a quantitative analysis on the incidence of preparation errors in the robot groups and manual groups. The data could not be merged because of the different medicine types in each study, but the results of all studies showed that the incidence of preparation errors in the robot groups ranged from 0.0% to 0.9%, which was significantly lower than that in the manual group (0.013%–12.5%) [13-15,17,19,26,31]. One study showed that the number of daily preparation errors was reduced from 0.26 ± 0.78 to 0.06 ± 0.13 . One non-controlled study  and one in-depth interview  also showed that the robots could reduce the incidence of preparation errors by warning, without interference by factors such as visual fatigue. Finally, a 3-year real-world study  showed that the percentage of actual failed preparations was <1% (0.21%, 0.15%, and 0.18% in 2015, 2016, and 2017, respectively).
Three studies [9,10,42] evaluated the incidence of preparation failures, including one controlled study and two non-controlled studies. The model used in two of the studies was the CytoCare, and the incidence of failures was 12% (n = 1028)  and 19% (n = 4509) , respectively. The model used in another study was the APOTECAChemo, and the incidence of failures was 2% (n = 11,642) . The reasons for preparation failures included robot shutdowns, mechanical failures, re-calibration, and other practical problems . However, as the application time increases, the rate of preparation success might gradually improve. For example, one study  showed that the rate of preparation success increased from 76.8% in week 1 to 95.3% in week 10.
Two RCTs [22,25] evaluated the properties of the solutions prepared by the robots. In one study , measurement of the size, pH value, and osmotic concentration of lipid particles in the nutrient solutions showed that the nutrient solutions prepared by the automated compounding device were superior to those prepared by gravity infusion in terms of stability and compatibility. By measuring antibody aggregation, another study  concluded that robotic compounding of monoclonal antibodies was feasible and that the robot could be used to achieve reproducible high-quality compounding for delicate formulations.
One non-controlled study  evaluated the robots’ role in error recognition. The study showed that the robot recognized 1.12% (n = 133) of errors that would have caused harm to patients.
Safety: Twenty-two studies evaluated safety (Table 3), including 13 controlled studies, 8 non-controlled studies, and 1 qualitative study. The evaluation indicators were product pollution, environmental pollution, and health damage to technical personnel.
Eleven studies [27,21,39,30,45,43,41,38,36,35,46] evaluated product pollution, including three controlled studies, seven non-controlled studies, and one qualitative study. Five studies [39,30,41,38,36,35] did not detect microbial pollution, and one non-controlled study  showed that non-contaminated bags were not contaminated after preparation, revealing that the robot enabled the preparations with low levels of product contamination. One controlled study  showed that the positive rate of bacterial cultures in the piston decreased from 26.67% to 13.33%, revealing that the robot could significantly reduce the possibility of product pollution. One in-depth interview  revealed that the robot was cleaner than the biological safety cabinet because it had an air channel cleaning system and a closed negative-pressure environment. However, another study  showed that the microbial culture positive rate was 7.9% (n = 152), which was related to product pollution. Another investigation showed that microbial contamination might originate from the washing station, which was easily be ignored . Finally, the technicians’ gloves were also key sites of microbial contamination .
Eight studies [24,23,39,18,43,38,36,35] evaluated environmental pollution, including three controlled studies and five non-controlled studies. Two controlled studies [24,18] showed that the external pollution of the robots was relatively low through monitoring gloves, infusion bags, and other equipment, indicating that environmental pollution could be improved. One non-controlled study  detected only one type of contamination associated with pulling the needle out of the syringe, which was very small (spots of <3 mm) but relatively frequent (11%). Pollution was mainly observed inside the robot , especially in the area of the dosing device [24,23,39]. The robot was able to perform automatic microbial decontamination by ultraviolet radiation . One controlled study and one non-controlled study showed that the settling plate/contact plate met the recommended limits for the class A area of the clean room  and that the surface and air samples complied with an ISO 5 class environment , indicating that the robots had well-controlled programs.
Eleven studies [26,17,23,15,16,14,13,19,43,40,46] assessed health damage to the technical personnel, including eight controlled studies, two non-controlled studies, and one qualitative study. Seven controlled studies [26,17,15,16,14,13,19] performed a quantitative analysis of the incidence of health damage to technical personnel in both the robot groups (chapped fingers: n = 0, ampoule scratches: n = 0, fatigue: n = 0–1, syringe stabs: n = 0–3) and manual groups (chapped fingers: n = 1–29, ampoule scratches: n = 2–38, fatigue: n = 3–19, syringe stabs: n = 1–28). The results showed that the incidence of health damage to technical personnel was significantly lower in the robot groups (0.0%–2.9%) than in the manual groups (5.1%–40.0%), and the somatic pain score was also significantly lower in the robot groups than manual groups (2.65 ± 0.47 vs. 5.76 ± 0.03, respectively). One non-controlled study  also showed that the robot could provide effective occupational protection for nurses. In an in-depth interview , all interviewees were nurses in the oncology ward and daytime chemotherapy center, and they had strong experience with occupational protection provided by the robots. Two studies [23,43] showed that no contaminant exposure was found in the gloves or on the hands of the technicians by applying the robots.
Economy: Six studies [34,26,10,7,30,8] conducted an economic evaluation, including four controlled studies, one non-controlled study, and one economic study. The evaluation indicators were labor costs and material costs. Through 1000 simulations, an economic study  showed that robots could reduce healthcare costs by preventing medication errors and reducing adverse drug events, saving an average of $288,350 per year. Three controlled studies [26,7,30] showed that robots could reduce costs by 8% to 66% by saving materials, and the cost savings might continue to rise with the increase in preparations . One controlled study  showed that a robot could reduce the labor costs of three pharmacists, but two other studies [26,10] showed that the robots could not reduce labor costs, could not improve the full-time equivalent of a hospital general pharmacist/technician, and even needed additional on-site engineers to help resolve possible breakdown.
Social suitability: Only one study  investigated technicians’ satisfaction with robots. The results showed that responders agreed that the overall impression of robots was “very good,” and the safety features of robots had a median score of being “very safe”.