In this prospective study, 173 patients suspected of having BSIs were enrolled from October 2022 to October 2023 ( Figure 1 ). The diagnosis of suspected BSI was made through the clinical judgment of the treating physicians according to systemic inflammatory response syndrome/sepsis criteria (Evans et al., 2021). The exclusion criteria included age < 18 years, having a mental disorder, being pregnant, or having any terminal-stage disease. Upon the suspicion of a BSI, whole blood samples were obtained for the BC and ddPCR assay simultaneously. Clinical and laboratory data were also collected.

This study was approved by the institutional review board of the First Affiliated Hospital of Sun Yat-sen University. Written informed consent was obtained from all patients or their legal representatives. All data were anonymized prior to analysis.

For each patient with a suspected BSI, at least two sets of BCs were collected according to routine clinical practice (BacT/ALERT^® VIRTUO System, bioMérieux, France) (Murray and Masur, 2012). Once the system showed a positive signal, gram staining and subculture were performed. Following overnight incubation, the pathogens were identified by matrix-assisted laser desorption/ionization time-of-flight mass spectrometry (VITEK^® MS system, bioMérieux, France). AST was performed using the VITEK^® system (bioMérieux, France). Interpretation of the AST results was based on the Clinical and Laboratory Standards Institute guidelines (CLSI, 2022). For Enterobacterales with a carbapenem-resistant phenotype, carbapenem-encoding resistance genes (bla _KPC, bla _NDM, bla _VIM, bla _IMP, and bla _OXA48) were detected using a Xpert Carba-R Assay (Cepheid, USA) following the manufacturer’s instructions (Jin et al., 2021b). For most of Staphylococcus spp., cefoxitin was used to predict results for mecA-mediated methicillin (oxacillin) resistance (CLSI, 2022). If gram-positive (G⁺) strains were shown to be the vancomycin-resistant phenotype by AST, PCRs were performed to detect vancomycin-resistant genes as described previously (Selim, 2022).

For ddPCR testing, at least 5 mL of whole blood using EDTA anticoagulant was collected. After centrifuging at 1200×g for 5 min, 2 mL of plasma was collected for DNA extraction using the Easy-CF2 Nucleic Acid Extraction/Purification Kit (Pilot Gene Technologies, Hangzhou, China) according to the manufacturer’s instructions.

Further, ddPCR testing was conducted as previously described with some modifications via a multiplex ddPCR testing platform for research use only (Pilot Gene Technologies, Hangzhou, China) (Wu et al., 2022). Briefly, the reaction mixture containing 15 μL of DNA extract was passed through a micro-channel (Droplet Generator DG32) to generate tens of thousands of water-in-oil emulsion droplets. After PCR amplification using a Thermal Cycler TC1, droplet counts and amplitudes were scanned and analyzed by a CS7 chip scanner and Gene PMS software (V1.1.8.20221121). Copies of the targeted pathogens or genes were reported. The synthesized DNA fragment at a concentration of 10⁴ copies/mL was used as the positive control, which was prepared by inserting pathogen target DNA into the pUC57 plasmid. DNase-free water was used as the negative control. The testing process took < 2.5h in total.

Table 1 presents the detection of pathogens and antimicrobial resistance (AMR) genes with the corresponding fluorescence channels. Considering the global and local pathogen epidemiological BC data (Jin et al., 2021a; Chen et al., 2023; Zhang et al., 2024), target pathogens included 12 common BSI bacteria, three common BSI fungi, and five main resistance markers. In addition, cases of clinical diagnosis of aspergillosis, mucormycosis, and Pneumocystis jirovecii pneumonia assisted by blood-based next-generation sequencing have been increasing in recent years (Cheng and Yu, 2022; Hoenigl et al., 2023; Wang et al., 2024), so the corresponding pathogens were also recommended for detection by clinicians in this study ( Table 1 ).

The detection sensitivity was determined according to the limit of detection (LoD) which was defined as the lowest amount of a target that is consistently detectable in at least 95% of the samples tested (Canchola et al., 2019). In this platform, LoD was determined by a series of the manufacturer’s experiments, of which two-fold serial dilutions of microbial DNA were taken. For each dilution of a pathogen, the lowest concentration was used as the LoD of the pathogen when at least 19 of 20 replicates could be detected. According to the manufacturer’s instructions, the detection sensitivity of this platform was 50 copies/mL for coagulase-negative Staphylococci (CoNS), Streptococcus spp., and Candida spp., and 25 copies/mL for the other microorganisms.

The results of the ddPCRs and BCs were compared among the 173 samples. A positive BC (BC+) indicated a positive result in the BC, which excluded the possibility of contamination by clinical assessment. A positive ddPCR result (ddPCR+) indicated the presence of at least one target pathogen in ddPCR testing, while a negative result (ddPCR-) indicated the absence of any target pathogen. For ddPCR+/BC+ samples, pathogen results were further compared between the two methods. If the pathogen(s) via ddPCR testing displayed 100% or 0% concordance with BC, consistency or inconsistency was considered respectively; otherwise, partial consistency was considered. In addition, laboratory turnaround time (LTAT) refers to the time from receipt of the sample in the laboratory to the reporting of results (Of and For, 2014).

The results of the ddPCR were available to clinicians in real-time. Clinical infection and the outcome of these patients were verified by two trained physicians independently. A culture-proven BSI was defined as a patient’s samples being BC+ with systemic signs of infection, which may be secondary to a documented source or primary, according to the definitions released by the National Healthcare Safety Network (Timsit et al., 2020). If infection was suspected, all microbiological examinations were collected within 7 days of enrolment according to the standard microbiology laboratory procedures, which included routine microbiological cultures, microscopic examination, and other examinations such as metagenomic next-generation sequencing. A composite clinical infection standard was defined, consisting of all microbiological results and clinical adjudication (Kalligeros et al., 2020; Nguyen et al., 2019).

For ddPCR+/BC- results, the following classifications were used according to previous studies (Kalligeros et al., 2020; Nguyen et al., 2019): (i) probable: ddPCR result was concordant with a microbiological test performed within 7 days of sample collection from another extra-blood site; (ii) possible: without microbiological data but the ddPCR result had the potential for pathogenicity based on clinical presentation and laboratory findings; (iii) putative false: ddPCR result was inconsistent with clinical presentation.

Statistical analysis was performed using SAS 9.4, using Chi-squared test, analysis of variance, Student’s t-test, Wilcoxon Rank-Sum test, or logistic regression analysis, where applicable.

In addition, receiver operating characteristic (ROC) curves were constructed, and the areas under the ROC curves (AUCs) were estimated to determine the feasibility of using the DNA load of a pathogen as a predictor for a corresponding BSI. Furthermore, Cohen’s kappa coefficient of agreement was calculated, which was used to classify the level of concordance: poor (<0.00), slight (0.00-0.20), fair (0.21-0.40), moderate (0.41-0.60), substantial (0.61-0.80), almost perfect (0.81-1.00) (Landis and Koch, 1977).

A P value of < 0.05 was considered statistically significant.

A total of 173 samples were obtained from the 173 patients for both BC and ddPCR testing. Patient characteristics are presented in Table 2 . The average age of the patients was 63.9 years (range: 18 - 102), and 64.7% (112/173) were men. In total, 65.9% (114/173) of the patients were in the ICU.

As shown in Figure 1 , microorganisms were detected in 94 samples (54.3%) by BC and/or ddPCR testing; of these, 92 were ddPCR+. Compared to BC-/ddPCR- patients, BC+ and/or ddPCR+ patients showed significantly higher levels of procalcitonin (PCT) (P = 0.001), while the prevalence of pulmonary infection (P = 0.009), and recent use of broad-spectrum antibiotics (P = 0.034) were significantly lower ( Table 2 ).

The results of BC and ddPCR were concordant in 125 samples, including 79 BC-/ddPCR- and 46 BC+/ddPCR+ samples ( Figure 1 ). For the 46 BC+/ddPCR+ samples, the pathogen results were completely consistent in 18 samples, partially consistent in 21 samples, and inconsistent in 7 samples ( Figure 1 ). When taking the pathogen results into consideration, the rate of samples with completely consistent results was 56.1% (97/173). The inconsistent results of seven BC+/ddPCR+ samples are listed in Supplementary Table S1 .

For the 21 samples with partially concordant results, all the culture-proven pathogens were detected by ddPCR simultaneously. The LTAT was further analyzed among the BC+/ddPCR+ samples, excluding those with inconsistent results (n = 39) ( Figure 2A ). For the BCs, the average reporting time of microscopy and isolate identification was 22.0 h (6.9 - 44.2 h) and 45.1 h (17.4 - 103.0 h) respectively, much longer than that of ddPCR testing (5.4 h, 2 -7.8 h) (P < 0.0001).

In addition, there were two BC+/ddPCR- samples due to the microorganisms being outside the detection range of ddPCR ( Supplementary Table S1 ). Of the 46 BC-/ddPCR+ samples combining microbiological and clinical evidence, 21 (45.7%) met the criteria for a probable BSI, 11 (23.9%) for a possible BSI, and the remaining 14 cases (30.4%) were putatively false ( Figure 1 ).

Table 3 lists the overall agreement between the BC and ddPCR results. The aggregate sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) for ddPCR were 81.3%, 63.2%, 45.9%, and 89.8%, respectively. Compared to G⁺ (80.0%) and G^- bacteria (87.2%), the sensitivity of ddPCR was much lower for fungi (40.0%). If both probable and possible BSIs were assumed to be positive, the anticipated sensitivity, specificity, PPV, and NPV of ddPCR for complicated BSIs were 88.8%, 86.0%, 84.5%, and 89.9%, respectively.

Culture-proven BSIs were positive for 52 pathogens in 48 (27.7%) cases; a polymicrobial BSI was detected in 8.3% of cases (4/48) ( Figures 1 , 2B ). Among these pathogens, 42 (80.8%) were G^- bacteria; the three top strains were Escherichia coli (36.5%, 19/52), Klebsiella pneumoniae (21.2%, 11/52) and Pseudomonas aeruginosa (7.7%, 4/52) ( Figure 2C ). As shown in Figure 2C , 18 types of pathogens were detected by BC, four of which were outside the detection range of ddPCR, including Burkholderia multivorans (n = 1), Salmonella typhimurium (n = 1), Pandoraea spp. (n = 1), and C. auris (n = 1).

Additionally, four bla _KPC genes were found in carbapenem-resistant K. pneumoniae strains, which were detected by ddPCR simultaneously ( Figure 2C ). None of the carbapenem-resistant Enterobacterales strains expressed AMR genes of bla _NDM, bla _VIM, bla _IMP, or bla _OXA48. For Staphylococcus spp., none were methicillin-resistant. None of the G⁺ strains were vancomycin-resistant.

Among the 92 ddPCR+ cases, 53 (57.6%) had a single microorganism detected, while there were multiple microorganisms (2 - 4 pathogens) detected in the other 39 (42.4%) ( Figure 2B ). A total of 143 microorganisms were detected, comprising G^- strains (n = 102), G⁺ strains (n = 23), and fungal strains (n = 18); of them, 41 (28.7%) were detected by BC simultaneously. Within the detection range of ddPCR, 7 culture-proven pathogens were not found in the ddPCR testing ( Figure 2C ).

E. coli (21.0%, 30/143) was the most prevalent pathogen, followed by P. aeruginosa (17.5%, 25/143), K. pneumoniae (14.7%, 21/143), and Acinetobacter baumannii (11.2%, 16/143). All the carbapenem-resistant genes, bla _KPC (n = 8) and bla _NDM (n = 1), were detected from K. pneumoniae strains. For G⁺ microorganisms, Enterococcus spp. (9.1%, 13/143) was the most prevalent strain. Among the microorganisms from the BC-/ddPCR+ samples, the vanA/vanM genes were found in three strains of Enterococcus spp., and one CoNS strain was positive for the MecA gene in ddPCR testing. Besides Candida spp. (n = 4), the remaining 14 strains of fungi were Aspergillus spp. (n = 7), Pneumocystis jirovecii (n = 5), and Cryptococcus spp. (n = 2), respectively.

The DNA loads of the microorganisms detected by ddPCR were further analyzed, showing a median level of 158.0 (range: 30.0 - 3.2 × 10⁵) copies/mL. In total, 72.7% (104/143) of the microorganisms had DNA loads < 1,000 copies/mL ( Figure 3A ). Among the 41 microorganisms detected by both methods, the rate of microorganisms with low levels (< 1,000 copies/mL) was 43.9% (18/41), much lower than the rate of those detected by ddPCR only (84.4%, 86/102) (P < 0.0001) ( Figure 3B ).

Of the top five microorganisms detected by ddPCR, 50.0% (15/30) of E. coli and 52.4% (11/21) of K. pneumoniae were found to have levels > 1,000 copies/mL, which were more than A. baumannii (37.5%, 6/16) and Enterococcus spp. (23.1%, 3/13) ( Figure 3C ). The detection rate of E. coli by BC was found to increase with increasing DNA loads (odds ratio = 1.002, 95% confidence interval = 1.001-1.002, P = 0.0002) ( Figure 3D ). A similar trend was also found for K. pneumoniae (odds ratio = 1.001, 95% confidence interval = 1.000-1.002, P = 0.037) ( Figure 3D ). Surprisingly, the DNA loads of P. aeruginosa were all below 500 copies/mL, 44.0% (11/25) of which were < 100 copies/mL ( Figure 3C ).

Among the two A. baumannii strains found by BC, only one had detected by ddPCR simultaneously with a DNA load of 49,427.0 copies/mL. Enterococcus spp. was detected in one sample by both methods with a DNA load of 301.5 copies/mL. The detection rate of Staphylococcus aureus by ddPCR was only 1.4% (2/143), with extremely low levels (< 35 copies/mL).

Taking BC as the gold standard, the ROC curve was used to reveal whether the pathogen load could be a potential marker for a corresponding BSI for the top five detected microorganisms by ddPCR ( Figure 4 ).

The DNA load of E. coli discriminated an E. coli BSI from a suspected BSI with an AUC of 0.954 (95% CI: 0.898-1.000). At the cut-off value of 93.0 copies/mL, the sensitivity and specificity for E. coli levels were 94.7% and 93.3%, respectively. For K. pneumoniae BSI, the DNA load of K. pneumoniae at the cut-off value of 196.5 copies/mL yielded an AUC of 0.994 (95% CI: 0.986-1.000) with 100.0% sensitivity and 97.5% specificity. The cut-off value for P. aeruginosa was 36.0 copies/mL with 75.0% sensitivity and 87.0% specificity (AUC: 0.816, 95% CI: 0.560-1.000). Though the AUC was recorded as 0.728, the 95% CI for A. baumannii was 0.195-1.000 (P = 0.402). Remarkably, the AUC for Enterococcus spp. was 0.282 (95% CI: 0.000-0.778).

According to the ROC analyses above, we further performed a consistency analysis of the BC and ddPCR results ( Table 4 ). The Kappa coefficient showed almost perfect and good agreement for K. pneumoniae (κ = 0.834) and E. coli (κ = 0.731) BSIs, respectively. Regrettably, there was a slight agreement for P. aeruginosa BSIs when comparing ddPCR to BC (κ = 0.167).

According to their clinical presentation and all the laboratory findings, the antibiotic treatment regimen was adjusted for a patient if a composite clinical infection was diagnosed by the treating physician.

E. coli and/or K. pneumoniae were detected in eight BC-/ddPCR+ samples with DNA loads above the corresponding cut-off values. For other samples that were BC-/ddPCR+, pathogens with DNA loads of > 1,000 copies/mL were detected in seven, including A. baumannii (n = 3), Enterococcus spp. (n = 2), and Pneumocystis jirovecii (n = 2). As presented in Supplementary Table S2 , for these aforementioned BC-/ddPCR+ patients, the infection symptoms obviously improved when they received correspondingly adjusted antibiotic treatments.