Development and validation of LPCPIS: an integrated clinical, biomarker and lung ultrasound score for the diagnosis of ventilator-associated pneumonia

Abstract

Objectives

To evaluate whether the performance of the modified Clinical Pulmonary Infection Score (CPIS) for diagnosis of ventilator-associated pneumonia (VAP) could be improved by incorporating procalcitonin (PCT), lactate and a standardised lung ultrasound score (LUS).

Design

Prospective observational study.

Setting

Internal medicine and respiratory intensive care units (ICUs) of a tertiary care centre in India over 18 months.

Participants

97 adult patients receiving invasive mechanical ventilation for 48 hours or more with suspected VAP.

Primary and secondary outcome measures

The primary outcome was the diagnostic accuracy (area under the receiver operating characteristic curve (AUC)) of the integrated score compared with a reference standard of quantitative bronchoalveolar lavage culture (>10⁴ CFU/mL). Secondary outcomes included ICU length of stay and 14-day mortality.

Results

Of the 97 patients, 70 (72.2%) were culture positive. The modified CPIS alone demonstrated limited diagnostic utility (AUC 0.686). In contrast, the novel 4-point Lactate-Lung Ultrasound-Procalcitonin-Clinical Pulmonary Infection Score (LPCPIS) score (integrating CPIS >5, lactate >2.15 mmol/L, PCT >1.3 ng/mL and LUS >14) achieved excellent discriminatory ability (AUC 0.919, 95% CI 0.844 to 0.976). At a cut-off of ≥2, the score yielded 95.7% sensitivity and 77.8% specificity. Crucially, a 4-point LPCPIS bedside score of ≤1 was associated with a highly diminished probability of VAP, yielding a negative predictive value of 100% in this cohort. In comparison, a score ≥3 strongly indicated a high likelihood of VAP, with a positive predictive value of 96.5% and specificity of 92.6%, pending definitive microbiological confirmation.

Conclusions

The novel 4-point LPCPIS bedside score demonstrates significant diagnostic superiority over the traditional modified CPIS (AUC 0.919 vs 0.686). By integrating bedside lung ultrasound with biomarkers, this simplified pragmatic tool provides impressive discriminatory power, allowing clinicians in resource-limited settings to reliably assess probability of VAP and guide early antimicrobial decision-making while awaiting definitive microbiological cultures.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• Ventilator-associated pneumonia (VAP) is difficult to diagnose due to the poor specificity of clinical signs, and the traditional Clinical Pulmonary Infection Score (CPIS) has demonstrated limited diagnostic accuracy.

WHAT THIS STUDY ADDS

• This study derives a novel 4-point Lactate-Lung Ultrasound-Procalcitonin-Clinical Pulmonary Infection Score (LPCPIS) bedside score integrating lung ultrasound, procalcitonin and lactate with clinical signs, which achieved an area under the receiver operating characteristic curve of 0.919 and a negative predictive value of 100% for identifying patients with a low probability of VAP.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• The LPCPIS score offers a rapid, non-invasive ‘rule-out’ tool for resource-limited intensive care units, potentially reducing unnecessary empirical antibiotic use while awaiting culture results.

Introduction

Ventilator-associated pneumonia (VAP) is a pneumonia that develops after 48 hours of initiation of invasive mechanical ventilation (IMV) and remains the most common serious intensive care unit (ICU)-acquired infection globally.¹ India reports some of the highest VAP rates globally (8.9–46/1000 ventilator days).^{2 3} No definitive gold standard exists for the diagnosis of VAP.⁴ Clinical signs, such as fever, leucocytosis, infiltrates or purulent secretions, are non-specific and common in critically ill patients, frequently overlapping with non-infectious syndromes like atelectasis, pulmonary oedema, acute respiratory distress syndrome (ARDS) or systemic inflammation.4,6 These definitions have poor sensitivity, missing up to 60–70% of cases and thus have limited real-time clinical utility.^{7 8} Clinical Pulmonary Infection Score (CPIS) was introduced in 1991 by Pugin et al to standardise bedside diagnosis using six clinical, radiological and microbiological parameters (reported sensitivity of 93% and specificity of 100%).⁹ However, this score was developed and validated in a cohort of 28 patients and 40 pairs of samples, and this performance has not been replicated. A 2004 study showed 83% sensitivity, 17% specificity and an area under the receiver operating characteristic curve (AUC) of 0.55.¹⁰ A subsequent meta-analysis by Shan et al¹¹ confirmed the modest performance and revealed a pooled sensitivity of 65% and specificity of 64% (AUC 0.748).¹¹ Research has since then shifted towards the integration of imaging and biomarkers along with clinical data. A 2014 meta-analysis addressing this found that elevated procalcitonin (PCT) was a strong predictor of VAP (diagnostic OR (DOR) 17.9), far superior to CPIS (DOR 4.86).¹² Subpleural consolidations and dynamic air bronchograms on lung ultrasound strongly indicate pneumonia and have been quantified using the lung ultrasound score (LUS).^{13 14} While few attempts have been made to combine these elements, such as the paediatric CPIS-PLUS (LUS and PCT) or the (Chest Echography and Procalcitonin Pulmonary Infection Score (CEPPIS), utilising LUS and PCT), no published research to date has specifically tested a composite score that includes components from CPIS, PCT and lactate in an adult population.^{15 16} This study aimed to evaluate whether adding PCT and lactate to modified CPIS improves diagnostic accuracy and to explore the usefulness of an additive bedside score incorporating LUS.

Materials and methods

Study design and setting

This prospective observational study was conducted in the internal medicine and respiratory ICUs of a tertiary care centre over 18 months, following the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines. Adult patients (>18 years) receiving IMV for 48 hours or longer were screened. Inclusion required clinical suspicion of VAP,¹⁷ defined as new/worsening radiographic infiltrates plus ≥2 of fever (≥38.5°C) or hypothermia (<36°C); leucocytosis (>10⁹/L) or leucopenia (<4000/µL); PaO₂/FiO₂ <300; or purulent secretions. Patients with prior VAP or confounding inflammation sources (eg, ischaemia, trauma, burns) were excluded.

Patients were followed from the initial suspicion until discharge from the ICU. Baseline demographics, comorbidities, antibiotic exposure and severity scores (Sequential Organ Failure Assessment (SOFA), Acute Physiology and Chronic Health Evaluation II (APACHE II)) were recorded. The primary outcome was the diagnostic accuracy (AUC) of the integrated score compared with the reference standard. Secondary outcomes included 14-day mortality and ICU length of stay (LOS).

Reference standard

VAP was defined by a positive quantitative culture (>10⁴ CFU/mL) from distal airway sampling (Bronchoalveolar lavage (BAL), or mini-BAL), processed via automated systems (VITEK 2, MicroScan) as per Clinical and Laboratory Standards Institute (CLSI) guidelines. The CPIS score was adapted from the original criteria by Pugin et al (1991).⁹ To allow for real-time diagnostic evaluation before microbiology results are available, a modified CPIS (referred to as ‘CPIS alone’ or ‘baseline CPIS’ in the results) was calculated, which comprises:

• Temperature (Celsius): 0 points (≤38.4 and ≥36.5), 1 point (≥38.5), 2 points (≥39 or ≤36).

• Leucocyte count (x10⁹/L): 0 points (≥4 or ≤11 x 10⁹/L), 1 point (<4 or >11 x 10⁹/L).

• Oxygenation PaO₂/FiO₂: 0 points (>240 or presence of ARDS), 2 points (≤240 or absence of ARDS).

• Tracheal secretion: 0 points (absent or minimal secretions), 1 point (non-purulent), 2 points (purulent).

• Chest radiograph: 0 points (no infiltrate), 1 point (patchy infiltrate), 2 points (localised infiltrate).

For our analysis, the microbiological component of the original Pugin et al (1991) score was excluded to prevent incorporation bias. For PCT, venous blood samples were collected and analysed using a YHLO iFlash 1800 Chemiluminescence Immunoassay Analyzer. Venous or arterial blood samples were collected in heparinised syringes and analysed within 15 min of withdrawal using Radiometer ABL800 FLEX blood gas analyser for lactate measurements. As part of established standard of care of using point-of-care lung ultrasound, a 12-zone quantitative LUS was performed using a curved 2–5 MHz probe (SonoSite M-Turbo) by qualified senior resident critical care trainees and fellows. To ensure advanced sonographic proficiency and reproducibility, all operators had completed a standardised 2-week formal training programme and possessed documented experience of performing at least 200 lung ultrasound examinations, and a systematic review of these sonographic assessments was done by the critical care attending faculty and the programme director. As LUS was performed by the primary treating team, operators were blinded to the biomarker panels and microbiological data, but not to the patient’s baseline clinical or laboratory data. A methodological limitation of this study is that formal interobserver agreement for LUS interpretation was not systematically calculated. The operators scanned six zones in each hemithorax (superior and inferior in anterior, posterior and lateral fields), and a composite score (range 0–36) was calculated by grading each zone from 0 to 3.¹⁸

• Score 0 for normal aeration (A lines with lung sliding).

• Score 1 for moderate loss of aeration (>3 B lines).

• Score 2 for severe loss of aeration (coalescent B lines or subpleural consolidation).

• Score 3 for complete loss of aeration (tissue-like consolidation pattern often with dynamic air bronchogram).

To minimise bias, laboratory personnel and statistical analysts were blinded to clinical data. Given the lack of previous similar studies analysing such an approach, a sample size of 97 was calculated (α=0.05, power=0.80) assuming a baseline diagnostic AUC of 0.6. Data were analysed using SPSS (V.23) and R.

The derivation of the predictive model was done through a multivariate logistic regression framework. To optimise the model for rapid bedside utility in resource-limited ICUs, continuous variables were dichotomised based on optimal discriminatory thresholds established via Youden’s J statistic. The log odds of culture-positive VAP were modelled as a linear combination of these categorical predictors using maximum likelihood estimation.

To prevent overestimation of disease probability in this high-prevalence cohort, calibration was evaluated using the Hosmer-Lemeshow goodness-of-fit test (p value ≥0.05 indicating adequate agreement between predicted and observed). Additionally, because predictive models evaluated on their derivation cohort inherently overestimate future performance, non-parametric bootstrapping (2000 resamples with replacement) was employed for internal validation to quantify mathematical overfitting (optimism) for the final intended clinical tool. Optimism was calculated as the difference between the diagnostic performance on the bootstrap sample and the performance on the original dataset, and these values were subtracted from the apparent diagnostic metrics to generate the optimism-corrected sensitivity, specificity, predictive values and AUC for the optimal diagnostic cut-off.

To evaluate the risk of reference standard bias secondary to the high prevalence of pre-enrolment antimicrobial administration (92.8%), a stratified sensitivity analysis was done—the diagnostic parameters of the 4-point Lactate-Lung Ultrasound-Procalcitonin-Clinical Pulmonary Infection Score (LPCPIS) bedside score (AUC, sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV)) were stratified by the duration of prior empirical antibiotic exposure (≥72 hours vs <72 hours), and DeLong’s test was applied to test for significance, with the null hypothesis stating no difference in diagnostic discrimination between the heavily treated and limited treatment subgroups.

Patient and public involvement

Patients or the public were not involved in the design, conduct, reporting or dissemination plans of this research.

Results

A total of 468 patients were admitted to the Medicine ICU and Respiratory Medicine ICU over the study duration. Of these, 102 patients met the full inclusion criteria for clinical suspicion of VAP, of whom five patients were excluded (three met exclusion criteria and two had critical missing data). Data were analysed for 97 patients, of whom 70 (72.2%) were confirmed to have VAP by the reference standard of a positive quantitative culture, and the remaining 27 patients (27.8%) turned out to be culture negative (figure 1). The mean age in our population was 51.41 (±16.20) years with a near-equal gender distribution (51.5% female, 48.5% male). The mean SOFA score was 8.94 (±3.39), and the mean APACHE II score was 20.01 (±6.71) on ICU admission, indicating a high burden of baseline illness. The mean duration of IMV before VAP suspicion was 8.61 (±6.52) days, indicating a predominantly late-onset VAP (>4 days) (table 1). Of the 97 patients with suspected VAP, 70 (72.16%) turned out to be culture positive. Most patients (n=90, 92.78%) were already receiving antibiotics for 72 hours or more at the time VAP was first suspected (table 1). The mortality burden in our population was high and did not differ significantly between the culture-positive and culture-negative groups. In the culture-positive group (confirmed VAP), 14-day mortality was 31.4%, increasing to 45.7% at 28 days and with an overall in-hospital mortality of 58.6% (table 1).

Figure 1. Study flow diagram. BAL, bronchoalveolar lavage; ICU, intensive care unit; IMV, invasive mechanical ventilation; VAP, ventilator-associated pneumonia.

Table 1. Baseline characteristics, clinical outcomes and component analysis of the study cohort (n=97).

Variable	Value/frequency
Demographics and severity
Age (mean±SD)	51.41 (±16.20) years
Gender (n, %)	50 (51.5%) female, 47 (48.5%) male
SOFA/APACHE II scores on admission (mean±SD)	8.94 (±3.39)/20.01 (±6.71)
Duration of ventilation (days)/prior antibiotics (%)	8.61 (±6.52) days/90 (92.8%)
Baseline lactate (mmol/L)/PCT (ng/mL) (mean±SD)	2.39 (±1.05)/8.28 (±12.87)
Comorbidities and habits
Comorbidities/dependences (descending frequency)	Hypertension (41.2%), diabetes mellitus (33%), respiratory (26.8%), COPD (22.7%), renal (18.6%), TB (14.4%), cardiac (11.3%), obesity (11.3%), neurological (10.3%), alcohol/smoking (27.5%), others (19.7%)
Outcomes: mortality
14-day/28-day/in-hospital mortality rates	31.4%/45.7%/58.6%
CPIS components: association with confirmed VAP	P value
Local signs: radiographic infiltrates/purulent secretions	<0.001/<0.001
Systemic signs: temperature/leucocytes/oxygenation (P/F)	0.655/0.249/0.180
LPCPIS risk stratification (14-day mortality)	Mortality rate
Score 0/score 1 (low risk)	25.0%/21.4%
Score 2/score 3 (moderate risk)	27.8%/23.1%
Score 4 (high-risk phenotype)	51.6%

This table summarises the baseline clinical characteristics and outcomes of the study cohort (n=97). The data highlight a high prevalence of prior antibiotic exposure (92.8%) and significant illness severity. The component analysis demonstrates that while local signs (radiographic infiltrates and purulent secretions) significantly predicted culture-positive VAP, systemic inflammatory markers included in the traditional CPIS (temperature, leucocytes, oxygenation) failed to distinguish between VAP and non-VAP cases. Additionally, stratification by the novel LPCPIS score identifies a distinct high-risk phenotype (score 4) associated with 51.6% 14-day mortality.

APACHE II, Acute Physiology and Chronic Health Evaluation II; COPD, chronic obstructive pulmonary disease; CPIS, Clinical Pulmonary Infection Score; LPCPIS, Lactate-Lung Ultrasound-Procalcitonin-Clinical Pulmonary Infection Score; PCT, procalcitonin; SOFA, Sequential Organ Failure Assessment; TB, tuberculosis; VAP, ventilator-associated pneumonia.

The diagnostic accuracy of the modified CPIS was modest, with an AUC of 0.686 (95% CI 0.564 to 0.808, p=0.005) (table 2). All individual biomarkers and integrated models demonstrated superior performance in comparison. The study met its primary objective: the integration of PCT and lactate with modified CPIS (CPIS+PCT+lactate) resulted in improved diagnostic accuracy, with an AUC of 0.896 (95% CI 0.812 to 0.981) (table 2). In an exploratory analysis, the weighted LPCPIS model—a multivariate logistic regression combining modified CPIS, PCT, lactate and LUS—achieved an AUC of 0.935 (95% CI 0.871 to 0.999) (table 2, figure 2A).

Table 2. Comprehensive diagnostic performance and logistic regression model derivation.

Variable/model	Sensitivity/specificity (%) (95% CI)	PPV/NPV (%) (95% CI)	LR+/ LR− (95% CI)	Diagnostic AUC (95% CI)	Multivariate aOR (95% CI)	Diagnostic threshold	β coefficient*	Assigned weight (point)
Individual variables
Modified CPIS	74.3 (62.4 to 83.9)/63.0 (42.4 to 80.6)	83.9 (72.3 to 91.4)/48.6 (30.8 to 66.7)	2.01 (1.20 to 3.34)/0.41 (0.25 to 0.67)	0.686 (0.564, 0.808)	4.91 (1.90 to 12.67)	>5	1.591	1
Lactate	75.7 (63.9 to 85.2)/70.4 (49.8 to 86.2)	86.9 (75.8 to 94.2)/52.8 (35.5 to 69.6)	2.56 (1.41 to 4.64)/0.35 (0.21 to 0.56)	0.730 (0.615, 0.846)	7.40 (2.75 to 19.93)	>2.15 mmol/L	2.001	1
PCT	87.1 (76.2 to 94.3)/77.8 (57.7 to 91.4)	91.0 (80.7 to 96.9)/70.0 (50.6 to 85.3)	3.92 (1.92 to 7.98)/0.17 (0.09 to 0.31)	0.825 (0.722, 0.927)	23.72 (7.54 to 74.61)	>1.3 ng/mL	3.166	1
LUS	80.0 (68.2 to 88.9)/77.8 (57.7 to 91.4)	90.3 (79.3 to 96.6)/60.0 (41.7 to 76.4)	3.60 (1.76 to 7.35)/0.26 (0.15 to 0.42)	0.789 (0.683, 0.895)	14.00 (4.76 to 41.22)	>14	2.639	1
Constant (intercept)	–	–	–	–		–	−3.456	–
Integrated models
CPIS+PCT	87.1 (76.2 to 94.3)/77.8 (57.7 to 91.4)	91.0 (80.7 to 96.9)/70.0 (50.6 to 85.3)	3.92 (1.92 to 7.98)/0.17 (0.09 to 0.31)	0.864 (0.766, 0.963)	–	–	–	–
CPIS+lactate	91.4 (82.3 to 96.8)/40.7 (22.4 to 61.2)	80.0 (69.6 to 88.1)/64.7 (38.3 to 85.8)	1.54 (1.18 to 2.01)/0.21 (0.10 to 0.47)	0.808 (0.716, 0.901)	–	–	–	–
CPIS+PCT+lactate	97.1 (90.1 to 99.7)/77.8 (57.7 to 91.4)	91.9 (83.2 to 96.7)/91.3 (72.0 to 98.9)	4.37 (2.15 to 8.89)/0.04 (0.01 to 0.15)	0.896 (0.812, 0.981)	–	–	–	–
CPIS+lactate+LUS	95.7 (88.0 to 99.1)/55.6 (35.3 to 74.5)	84.8 (75.0 to 91.9)/83.3 (58.6 to 96.4)	2.16 (1.41 to 3.29)/0.08 (0.03 to 0.23)	0.883 (0.807, 0.960)	–	–	–	–
Weighted LPCPIS model	90.0 (80.5 to 95.9)/77.8 (57.7 to 91.4)	91.3 (82.0 to 96.7)/75.0 (55.1 to 89.3)	4.05 (1.99 to 8.24)/0.13 (0.06 to 0.27)	0.935 (0.871, 0.999)	–	–	–	–
4-point LPCPIS score
Score ≥1	100.0 (94.9 to 100.0)/44.4 (25.5 to 64.7)	82.4 (72.6 to 89.8)/100.0 (73.5 to 100.0)	1.80 (1.29 to 2.51)/0.02 (0.001 to 0.26)†	–	–	–	–	–
Score ≥2	95.7 (88.0 to 99.1)/77.8 (57.7 to 91.4)	91.8 (83.0 to 96.9)/87.5 (67.6 to 97.3)	4.31 (2.12 to 8.73)/0.06 (0.02 to 0.17)	0.919 (0.844 to 0.976)	–	–	–	–
Score ≥3	78.6 (67.1 to 87.5)/92.6 (75.7 to 99.1)	96.5 (87.9 to 99.6)/62.5 (45.8 to 77.3)	10.61 (2.78 to 40.48)/0.23 (0.15 to 0.36)	–	–	–	–	–
Score ≥4	42.9 (31.1 to 55.3)/96.3 (81.0 to 99.9)	96.8 (83.3 to 99.9)/39.4 (27.6 to 52.2)	11.57 (1.66 to 80.84)/0.59 (0.48 to 0.73)	–	–	–	–	–

This table presents the derivation of the weighted logistic regression model, the translation to the simplified 4-point LPCPIS bedside score and the overall diagnostic accuracy and prognostic utility of the integrated score compared with baseline-modified CPIS (cut-off >5) and individual biomarkers. Multivariate analysis confirmed procalcitonin, lactate and lung ultrasound as strong independent predictors (aOR 23.72, 7.40 and 14.00, respectively). The simplified clinical model (score >2; AUC 0.919) resulted in minimal degradation of diagnostic accuracy compared with the optimal weighted model (AUC 0.935) and demonstrated clinically meaningful enhancement over baseline-modified CPIS alone (AUC 0.686).

^*Logistic regression coefficients.

^†A Haldane-Anscombe correction was applied to calculate the lower bound 95% CI for the negative likelihood ratio due to zero false negatives.

aOR, adjusted OR; AUC, area under the receiver operating characteristic curve; CPIS, Clinical Pulmonary Infection Score; LPCPIS, Lactate-Procalcitonin-Clinical Pulmonary Infection Score; LR, likelihood ratio; LUS, lung ultrasound score; NPV, negative predictive value; PCT, procalcitonin; PPV, positive predictive value.

Figure 2. Diagnostic performance and model calibration. (A) ROC curves demonstrating the superior discriminatory ability of the novel weighted LPCPIS-integrated models compared with individual biomarkers and baseline modified CPIS. (B) Calibration plot for the weighted LPCPIS model; risk deciles closely adhere to the ideal reference line, visually validating statistical congruence (Hosmer-Lemeshow p=0.852). CPIS, Clinical Pulmonary Infection Score; LPCPIS, CPIS integrated with lactate, procalcitonin and lung ultrasound score; LUS, lung ultrasound score; PCT, procalcitonin; ROC, receiver operating characteristic; VAP, ventilator-associated pneumonia.

To create a simplified, practical bedside tool, a multivariate logistic regression analysis was performed to confirm the independent predictive value of each of the four variables. The results showed that all four variables were strong independent predictors of a culture-positive VAP diagnosis (table 2). PCT was the strongest predictor (adjusted OR (aOR) 23.72), followed by LUS (aOR 14.00), lactate (aOR 7.40) and CPIS (aOR 4.91) (table 2). This indicates that each component of the weighted LPCPIS model provides unique, non-redundant diagnostic information. Based on the optimal cut-offs derived from the receiver operating characteristic analysis (Youden Index) for each of the four non-invasive variables, a simple, additive 4-point LPCPIS score was developed by awarding 1 point each for:

1. Modified CPIS >5.5 (or >5).

2. Lactate >2.15 mmol/L.

3. PCT >1.3 ng/mL.

4. LUS >14.5 (or >14).

As CPIS and LUS are discrete point scales, the cut-offs behave similarly for whole numbers (5 and 14, compared with 5.5 and 14.5 as identified by regression) in terms of diagnostic performance. Derived from these cut-offs, the simplified 4-point LPCPIS bedside score demonstrated excellent diagnostic discrimination and achieved an AUC of 0.919 (95% CI 0.844 to 0.976, p<0.001), retaining robust performance comparable to the theoretical weighted LPCPIS regression model.

The exploratory weighted multivariate logistic regression model demonstrated excellent overall discriminatory ability (AUC 0.935, 95% CI 0.871 to 0.999). Using maximum likelihood estimation, the derivation yielded the following exact predictive equation for the log odds of VAP (see table 2 for full disclosure):

$${\displaystyle {\displaystyle \begin{array}{l}\mathrm{logit}(P)={\beta }_0+1.591(\mathrm{C}\mathrm{P}\mathrm{I}\mathrm{S}\S gt;5)+2.001(\mathrm{L}\mathrm{a}\mathrm{c}\mathrm{t}\mathrm{a}\mathrm{t}\mathrm{e}\S gt;2.15)\\ +3.166(\mathrm{P}\mathrm{C}\mathrm{T}\S gt;1.3)+2.639(\mathrm{L}\mathrm{U}\mathrm{S}\S gt;14)\end{array}}}$$

The weighted LPCPIS model demonstrated exceptional calibration, mirroring the high underlying pretest probability of VAP in this cohort. The Hosmer-Lemeshow test revealed no significant deviation between predicted and observed VAP frequencies (χ²=1.976, df=5, p=0.852). This statistical congruence is visually validated in the calibration plot (figure 2B), where the plotted risk deciles adhere closely to the ideal reference line. When translating these complex fractional weights into the simplified 4-point LPCPIS bedside score—assigning a uniform weight of 1 point per crossed threshold to facilitate rapid clinical calculation—the diagnostic discrimination remained highly robust (AUC 0.919, 95% CI 0.844 to 0.976). Internal validation via 2000 bootstrap resamples confirmed that the optimism (overfitting bias) of this simplified, unweighted scoring system was negligible. At the optimal diagnostic threshold of ≥2, the optimism-corrected AUC was calculated at 0.919 (95% CI 0.844 to 0.976). The optimism-corrected diagnostic accuracy metrics showed minimal shrinkage: corrected sensitivity was 95.7%, specificity 77.8%, PPV 91.8% and NPV 87.5%. The near-zero deviation between the apparent and optimism-corrected metrics robustly validates the stability of the 4-point LPCPIS bedside score, confirming its reliability as an empirical decision support tool pending definitive microbiological culture (table 2).

Stratified sensitivity analysis for reference standard bias

The 4-point LPCPIS bedside score demonstrated impressive discriminatory power independent of antibiotic duration. The diagnostic AUC for patients with ≥72 hours of prior antibiotic exposure remained excellent at 0.921 (95% CI 0.842 to 0.980), compared with the limited-exposure subgroup (<72 hours, n=7, 0.915, 95% CI 0.810 to 0.990), and DeLong’s test revealed no significant variation in overall discrimination (p=0.784), confirming that the tool’s performance parameters are robust and not artificially inflated by pretest antimicrobial suppression, though this comparison is inherently underpowered by the small sample size in the limited-exposure arm.

This score provides a highly practical bedside approach to clinical decision-making in patients suspected of having developed VAP (table 2):

1. Low probability (score 0–1): a 4-point LPCPIS bedside score of 0 or 1 identified patients with a minimal likelihood of VAP, demonstrating an NPV of 100.0% in this population.

2. High probability (score ≥3): this threshold yielded a specificity of 92.1% and a PPV of 96.5%, acting as a high-confidence indicator for VAP prior to culture positivity.

3. ‘High Risk’ (score 4): this identified a high-mortality phenotype (51.6% at 14 days and 76% in-hospital), requiring immediate escalation of therapy and aggressive patient support (table 1).

To address the possible risk of overfitting as a result of developing and testing the score on the same population, an internal validation was performed by a bootstrap procedure, with 2000 resamples that rederived the cut-off points within each sample and calculated a corrected (adjusted) AUC for the 4-point LPCPIS. The original AUC (0.919) was adjusted to 0.915 (95% CI 0.862 to 0.982), with optimism-corrected sensitivity, specificity, PPV and NPV essentially unchanged from the apparent values. The minuscule difference between the original and corrected AUC indicates limited overfitting and establishes the robust discriminatory ability of the score within this dataset. Analysis of the individual components of CPIS was done to understand why it performs modestly. Only the local, observable signs—localised radiographic infiltrates (p<0.001) and purulent tracheal secretions (p<0.001)—were significantly associated with culture positivity, while systemic inflammatory markers (temperature p=0.655, leucocyte count p=0.249 and PaO₂/FiO₂ ratio p=0.180) did not show any significant association with VAP (table 1). The secondary objective was to assess the prognostic value of these scores for the prediction of mortality at 14 days and ICU LOS (table 2).

1. 14-day mortality: This objective was not met by the originally proposed integrated score. Modified CPIS alone, CPIS with Gram stain (CPIS-mic), CPIS with quantitative culture (CPIS-cult), PCT alone, lactate alone and even the integrated model (modified CPIS+lactate+PCT, AUC 0.612, p=0.074) all failed to demonstrate statistically significant prognostic ability. Only models that included LUS achieved modest statistical significance (eg, modified CPIS+lactate+LUS, AUC 0.638, p=0.028).

2. ICU LOS: This objective was also not met. A Spearman correlation analysis did not find any statistically significant association between any of the diagnostic modes and the total ICU LOS.

To facilitate immediate clinical translation and adherence to diagnostic principles, a step-by-step practical flow chart using the 4-point LPCPIS bedside score was made (figure 3B).

Figure 3. Pathophysiological framework and proposed clinical implementation of the LPCPIS. (A) The conceptual transition from the traditional CPIS relying on non-specific systemic markers to the new 4-point LPCPIS, which integrates clinical suspicion, bacterial load, metabolic stress and point-of-care imaging. (B) A step-by-step practical bedside algorithmic flow chart designed to guide targeted empirical antimicrobial therapy and support antimicrobial stewardship pending definitive microbiological cultures. CPIS, Clinical Pulmonary Infection Score; ICU, intensive care unit; LPCPIS, CPIS integrated with lactate, procalcitonin and lung ultrasound; LUS, Lung Ultrasound Score; PCT, procalcitonin; POC, point of care; VAP, ventilator-associated pneumonia.

Discussion

This study addresses the challenge of accurately diagnosing VAP in a high-resistance ICU setting to enable timely therapeutic interventions. The primary objective was to determine whether augmenting the modified CPIS with standard-of-care biomarkers, such as PCT and lactate, could improve its diagnostic accuracy. This objective was achieved—the integration of these biomarkers provided a statistically significant improvement in diagnostic power, creating a clinically useful tool. In our population, modified CPIS performed modestly for identifying culture-positive VAP when used in isolation (AUC 0.686, 95% CI 0.564 to 0.808) (table 2), which confirms the large body of evidence challenging the high sensitivity and specificity (93% and 100%, respectively) of the original 1991 Pugin et al study and aligns more closely with subsequent meta-analyses by Shan et al (AUC 0.748) and Schurink et al (AUC 0.55).9,11 This study contributes to the robust consensus by reaffirming that the CPIS, as originally constructed, is insufficient for clinical decision-making. Furthermore, our data demonstrate that this modest performance is due to the composition of the score. In a secondary analysis, we found that key markers of systemic inflammation—temperature (p=0.655), total leukocyte count (p=0.249) and the PaO₂/FiO₂ ratio (p=0.180)—did not differ significantly in relation to VAP. Only obvious local signs, such as purulent tracheal secretions (p<0.001) and radiographic infiltrates (p<0.001), were predictive, which aligns with the meta-analysis by Fernando et al,¹⁹ which revealed poor performance for individual signs, including fever (sensitivity 66.4%, specificity 53.9%).¹⁹ This score is diluted by ‘noise’ from systemic variables ubiquitous in critically ill patients (eg, atelectasis, pulmonary oedema), such that it is outperformed even by a single non-specific marker of metabolic stress—lactate (AUC 0.730) (table 2). In contrast, the CPIS+PCT model (AUC 0.864), the CPIS+PCT+lactate model (AUC 0.896) and the CPIS+PCT+lactate+LUS model (AUC 0.935) all achieved significantly higher diagnostic accuracies, suggesting a synergistic effect (table 2). Furthermore, we found that PCT alone was a strong independent predictor (OR 23.72), which expands on the meta-analysis by Sotillo-Díaz et al,¹² where elevated PCT was found to be superior to CPIS (DOR 17.9 vs 4.86) (table 2).¹² Our inclusion of lung ultrasound in this scoring system contributes to the growing body of evidence supporting its use in VAP diagnosis, as evidenced by its inclusion in the paediatric CPIS-PLUS (AUC 0.86) and CEPPIS scores (AUC 0.83).^{15 16} The central contribution of this study is the development of the 4-point LPCPIS derived from dichotomous cut-offs: CPIS >5, lactate >2.15 mmol/L, PCT >1.3 ng/mL and LUS >14, which achieved an AUC of 0.919 (95% CI 0.844 to 0.976). At the optimal Youden-derived cut-off of ≥2, the score showed a sensitivity of 95.7% and specificity of 77.8% (table 2). The LPCPIS integrates clinical suspicion, anatomical evidence, bacterial load and metabolic stress into a practical bedside tool for VAP diagnosis. In our population, a score of 4 identified a high-risk VAP phenotype, with 51.6% 14-day and 76% in-hospital mortality. We also found that no score shows a significant correlation with ICU LOS (Spearman’s p≈0.761). This suggests that while VAP increases short-term mortality, total ICU LOS is likely dictated by downstream complications of critical illness rather than a single clinical variable. Clinically, the ability to predict culture positivity or negativity prevents misdiagnosis of non-infectious mimics (eg, ARDS, atelectasis) and helps de-escalate antimicrobial therapy. VAP requires a multifaceted assessment, and future research should focus on refining synergistic models rather than searching for a single biomarker.

Interpretation

The study elucidates modified CPIS limitations and offers a practical alternative in the form of the 4-point LPCPIS bedside score, which uses four distinct, independent pillars of VAP pathophysiology (table 2, figure 3A):

1. PCT was the strongest predictor (aOR 23.72), consistent with prior studies.¹²

2. LUS was the second strongest predictor (aOR 14.00), confirming its utility as a reliable, non-invasive imaging tool.^{13 20}

3. Lactate (aOR 7.40) was validated as a novel diagnostic adjunct likely by detecting sepsis-induced tissue hypoperfusion that distinguishes VAP from colonisation.⁴

The scores and biomarkers that effectively diagnose VAP were statistically insignificant predictors of 14-day mortality and did not correlate with ICU LOS (table 2). This indicates that patient survival is determined not only by the mere diagnosis of culture-positive VAP but also by other factors of host response (ie, septic shock) and the resistance profile of the pathogen (multidrug-resistant organism), which were found to be the only strong independent predictors of mortality in separate analyses. This supports the low attributable mortality concept described by Bassi et al, where underlying severity drives outcomes rather than VAP itself.⁶ These findings require external validation.

Limitations

1. Limited generalisability (external validity): As a single-centre study with a unique microbial profile, generalisability is limited. While this cohort strongly reflects the epidemiology of tertiary Indian ICUs—characterised by high empirical antibiotic exposure and predominantly multidrug-resistant Gram-negative aetiology—these specific endemic microbial pressures necessitate further validation in future multicentre prospective studies across diverse intensive care settings with varying endemic microbial profiles and resistance patterns to ensure its broad generalisability as a robust diagnostic and antimicrobial stewardship instrument.

2. Reference standard: While both BAL and mini-BAL are acceptable methods for diagnosis, and our internal analysis did not reveal a statistically significant difference in diagnostic yield (p=0.503), this introduces a potential source of heterogeneity.

3. Potential reference standard bias: A significant proportion of our cohort (92.78%) received empirical antibiotics for ≥72 hours prior to enrolment. This introduces potential reference standard bias, as prior antimicrobial therapy can suppress culture positivity, potentially overestimating the score’s NPV. Furthermore, predictive values are highly dependent on the high baseline prevalence of VAP in this setting; likelihood ratios provide a more stable metric, but multicentre external validation is necessary to confirm generalisability across diverse microbial profiles.

4. A recognised methodological limitation of this study is that formal interobserver agreement (eg, kappa statistic) for LUS interpretation was not systematically calculated. However, this omission reflects a pragmatic study design and real-world utility. Point-of-care ultrasound is fundamentally defined as a dynamic, real-time extension of the physical examination.

Conclusion

The traditional CPIS is an insufficient diagnostic tool. A novel, 4-point LPCPIS score integrating clinical signs (modified CPIS) with markers of bacterial load (PCT), metabolic stress (lactate) and imaging (LUS) provides excellent diagnostic accuracy. This non-invasive, real-time bedside tool has the potential to guide therapy and help improve antimicrobial stewardship practices. While the specific cut-offs and the high AUC of LPCPIS require external validation, the conceptual framework of this study is highly generalisable. Future VAP research should abandon the search for a single ‘magic-bullet’ biomarker and instead focus on refining and validating multifaceted, synergistic models that integrate clinical assessment with markers of bacterial load, metabolic stress and imaging evidence.

Clinical significance

The 4-point LPCPIS score offers a practical, real-time algorithm for antimicrobial stewardship practice in high-resistance settings where culture delays (48–72 hours) are unavoidable.

1. Score 0–1 (low probability): this yields a high NPV in this population, supporting clinical confidence to safely de-escalate empirical antibiotics and actively pursue alternative, non-infectious diagnoses (eg, ARDS, atelectasis, oedema).

2. Score ≥3 (high probability): with a 96.5% PPV and 92.6% specificity, this strongly indicates a VAP diagnosis and justifies the continuation of targeted empirical antibiotic therapy as per the local antibiogram pending culture data.

3. Score 4 (high probability+high-risk phenotype): this identifies a high-risk phenotype with 51.6% 14-day mortality, indicating an immediate need for aggressive patient management and escalation of care.

Data availability statement

Data are available upon reasonable request.