Validation of The 4C Deterioration Model for COVID-19 in a UK Teaching Hospital During Wave 2.

Dear Editor

The 4C Deterioration model is a point-of-admission tool for predicting in-hospital clinical deterioration in patients with COVID-19. It was developed and validated using data from the UK first wave until August 26, 2020. Most point-of-admission models proposed for risk stratification in COVID-19 suffer from poor calibration. In contrast the 4C Deterioration model was shown to be well-calibrated and have good discriminative characteristics.

Since the model was developed, treatment of COVID-19 has evolved, including the use of corticosteroids as standard of care in hypoxemic respiratory failure, and new variants of the virus have emerged.3, 4, 5 Such changes can lead to declining model performance over time. Temporal validation during the second wave of the pandemic is therefore important to assess whether the discrimination and calibration of the model has been maintained.

Here we present the first external validation of the 4C Deterioration model using data from the UK second wave.

All adult patients admitted to Cambridge University Hospitals between August 27, 2020 and April 16, 2021 who tested positive for SARS-CoV-2 were included in the validation cohort. Diagnostic testing used either a real-time reverse transcription polymerase chain reaction (RT-PCR) of the RdRp gene from a nasopharyngeal swab, or the SAMBA II point-of-care test used at the hospital. Readmissions and inter-hospital transfers were excluded. Data were extracted from the electronic health record system (Epic) on May 14, 2021, meaning at least 28 days follow up were available for all included patients. The data were analyzed retrospectively in R 3.6.3.

All patients were treated as per detailed local guidance in use in the hospital at the time.

As in the development study, the outcome was in-hospital deterioration (commencement of ventilatory support, critical care admission, or death); cases tested more than seven days after admission were considered nosocomial; and patients who remained in hospital but had not deteriorated by the time of data extraction were classed as not deteriorating.

The 4C Deterioration model uses a combination of demographic factors (age and sex), blood tests (C-reactive protein, urea, lymphocytes), observations (respiratory rate, oxygen saturation (SpO2), Glasgow Coma Scale), requirement for supplemental oxygen, whether the infection was hospital-acquired, and the presence of lung infiltrates on radiographic chest imaging.

To calculate the risk score we used only results and observations recorded within 24 h of admission, or within 24 h of the time of first positive test for nosocomial cases. The development study accounted for missing values in their data using multiple imputation but did not report the imputation model parameters used, meaning that this approach cannot be used either in validation or in clinical practice. Missing values in the validation data were instead median imputed from the development dataset. This avoids the potential bias that would be introduced if only patients with a complete set of observations, blood results and imaging were included in the validation, as the presence or absence of observations or tests may reflect clinician assessment of the severity of disease.

To assess the discriminative performance of the proposed model we calculate the Area under the Receiver-Operating Curve (AUROC), where a value of 1 represents perfect discrimination and 0.5 discrimination no better than random chance. Additionally, we stratify the AUROC by month of patient admission, to investigate performance over time. We also calculate the Area under the Precision-Recall Curve (AUPRC; also known as the curve of positive predictive value (PPV) against sensitivity; true positives/(all positives) - true positives/(true positives + false negatives)), which measures discrimination relative to the observed incidence; and the number needed to evaluate (NNE, 1/PPV), defined as the number of patients predicted to deteriorate for every one additional correctly-detected deterioration, which is a measure of clinical burden. We assess model calibration using calibration-in-the-large and the calibration slope. We also visualize the calibration of the model through the observed incidence in each decile of predicted deterioration probability.

The study was approved by a UK Health Research Authority ethics committee (20/WM/0125). Patient consent was waived because the de-identified data presented here were collected during routine clinical practice; there was no requirement for informed consent.

The study included 950 patients. Compared to the development study, patients were younger (median 70 vs 75 years) and nosocomial infections were slightly more common (11.2% vs 9.9%). Other parameters were similar (Table 1 ). Missingness was greatest for urea (21.1% missing), radiology (15.9%), C-reactive protein (11.2%) and lymphocyte count (9.3%), all lower than in the development study.

TABLE 1

Patient baseline characteristics and parameters, outcomes and model performance.

	Our data	ISARIC 4C development study
Number of patients	950	74 944
Male, n (%)	498 (52.4%)	41 993 (56.1%)
Age, median [IQR]	70 [53, 82]	75 [60, 84]
Nosocomial infection, n (%)	106 (11.2%)	7320 (9.9%)
Glasgow Coma Scale, median [IQR]	15 [15, 15]	15 [15, 15]
Respiratory rate, breaths per minute, median [IQR]	19 [17, 23]	20 [18, 26]
Oxygen saturation, %, median [IQR]	96 [94, 97]	95 [92, 97]
Room air, n (%)	573 (60.3%)	48574 (69.4%)
Urea, mmol/L, median [IQR]	6.2 [4.5, 9.3]	7 [5, 11]
C-reactive protein, mg/L, median [IQR]	57 [22, 113]	80 [33, 154]
Lymphocyte count, x 109/L, median [IQR]	0.85 [0.59, 1.21]	0.9 [0.6, 1.3]
Radiographic infiltrates / number of patients with radiology result available	497/807 (61.6%)	29 579 / 47 749 (61.9%)
Outcomes*
Ventilatory support or critical care admission, n (%)	182 (19.2%)	15 039 (20.1%)
Death, n (%)	99 (10.4%)	16 885 (22.5%)
No deterioration, n (%)	669 (70.4%)	42 024 (56.1%)
Missing, n (%)	0 (0.0%)	996 (1.3%)
Model performance
AUROC [95% CI]	0.75 [0.71 to 0.78]	0.77 [0.76 to 0.78]
Calibration-in-the-large [95% CI]	-0.26 [-0.42 to -0.11]	0.00 [-0.05 to 0.05]
Calibration slope [95% CI]	1.00 [0.83 to 1.18]	0.96 [0.91 to 1.01]

Outcomes given here are the first point at which patients fulfil the composite endpoint for deterioration.

In-hospital deterioration occurred in 281 (29.6%) patients, compared to 42.6% in the development study. The lower risk of deterioration in this cohort may reflect differences in the patient population of the study hospital or could represent improvements in treatment over the course of the pandemic.

Figure 1 shows the performance metrics for the median-imputed data set. AUROC was 0.75 [95% CI 0.71 to 0.78]; calibration-in-the-large was -0.26 [-0.42 to -0.11], indicating overprediction of risk; and the calibration slope was 1.00 [0.83 to 1.18]. The NNE remains below 3.5 over the entire range of sensitivity, indicating that the clinical burden of use of the score is reasonable. The assessment of AUROC by month of admission (Supplementary Fig. 1) only revealed a slight decrease in the discriminative performance during the winter, when bed occupancy was reaching its peak.

Figure 1

(a) Histogram of predicted risk of clinical deterioration; (b) Receiver Operator Characteristic plot, with labels indicating the corresponding cutoff and the dashed line indicating the line of no discrimination; (c) Precision-Recall plot, with the 29.6% observed deterioration incidence indicated by the dashed line; (d) Calibration plot (with 95% CI), by tenths of predicted risk, with the dashed line indicating perfect calibration; (e) Number needed to evaluate (NNE) by sensitivity (recall). Abbreviations: AUROC, Area under the Receiver Operator Curve; TPR, true positive rate; FPR, false positive rate; AUPRC, Area under the Precision Recall Curve; PPV, positive predictive value.

Median imputation proved to be a viable approach to missing data, as the model's performance was not adversely affected by the imputed values. Additionally, this suggests that the model can be applied more widely, as many patients had missing values in at least one predictor: 384 (40.4%) patients in our study had at least one missing observation or result; compared to at least 36.3% (missing chest imaging alone; overall missingness unreported) in the development study. Performance was similar when patients with any missingness were excluded (AUROC 0.78 [0.74 to 0.82]; calibration-in-the-large -0.28 [-0.48 to -0.09]; calibration slope 1.09 [0.88 to 1.31]).

Despite slight overestimation of risk, discrimination and calibration remained consistent with the development study demonstrating robustness to the presence of novel variants and changes in treatment over time.

Declaration of Competing Interest

None declared.