Resources needed for facing the COVID-19 pandemic were dear, and multiple severity scores were suggested to predict mortality and maximize the benefit from the available resources. Albeit the higher need for them in developing and war-torn countries, these scores proved to be less feasible and showed poorer performance where resources are scarcer and when mortality is more dependent on the initial presentation of the patients. This study was successful in developing a new simplified resources-sparing scoring system for predicting COVID-19 mortality during hospitalization. LR-COMPAK relied on six variables easily obtainable within a short timeframe and still outperformed the previously available scoring systems in our sample. Not only did it aid in predicting mortality on an individual basis, but it also highlighted variations among cities, hospitals, and pandemic peaks. Moreover, its applicability extended to non-hospitalized patients, suggesting significant potential to aid in hospitalization decision-making and in the allocation of cases or resources across different centers. Finally, a customized version of the score, LR-ALBO-ICU, demonstrated notable effectiveness in forecasting mortality for ICU-admitted patients, potentially assisting in the allocation of available intensive care resources to those with higher predicted survival.
With a sensitivity of 87.9% and specificity of 72.1% (AUC of 0.88), our regression coefficient-based scoring system (LR-COMPAK) was able to explain more than 52% of the variability in mortality in our sample. This performance exceeded the discriminability of previously suggested scores including OuR-ARDSs (AUC of 0.85)34, CURB65 (AUC of 0.72)44, BURDEN (AUC of 0.83)35, NEWS2 score (AUC of 0.82)47, Chest Computed Tomography Severity Score (CTSS) (AUC of 0.84)48, qSOFA score (AUC of 0.74)43, as well as the score developed by Altschul et al. (AUC of 0.82)33. On the other hand, although other reported scores like CALL (AUC of 0.91)45, SOFA (AUC of 0.89)43, CMR (AUC of 0.90)49, and COVID-19 scoring system (CSS) (AUC of 0.92)50 performed slightly better in their own samples, they all included much more variables or at least one blood test or imaging study, which required incomparable longer time to acquire and rendered them inappropriate to aid decision making in our settings. Being simplified to only six easily accessible factors, namely: age, history of CKD, prior malignancies, HR, room blood oxygen saturation, and disturbed consciousness, the LR-COMPACK score could be obtained within 60 seconds even by non-medical personnel. This advantage extends also to several sites, where medical facilities did not afford the luxury of performing all required lab tests and imaging studies in the emergency room prior to admission51.
The variables of our developed LR-COMPAK score included one demographic characteristic (i.e., advanced age), which was one of the most reported risk factors for COVID-19 mortality52–54. Additionally, two comorbidities, namely malignancy and CKD, showed independent predictability of mortality in accordance with previous reports54–56 and were included in our score. The weakened immune system and the exaggerated susceptibility to more severe infections can explain the higher mortality in elderly57,58, as well as patients with malignancies59,60 or CKD61,62. On the other hand, LR-COMPAK also included three vital signs proved as independent predictors of mortality. Disturbed consciousness and tachycardia were among them, and albeit associating with multiple confounding factors for mortality prediction, they were reported among the risk factors of mortality in numerous studies63–68. Lastly, peripheral blood oxygen saturation was the individual parameter that carried the highest value in predicting mortality in our study, and this association was in line with other studies especially in countries with limited resources like Brazil, Peru, and South Africa 69–71. Hence, we utilized this correlation in our scoring system by categorizing the oxygenation percentage into eight bins with gradual mortality probabilities (appendix 1).
It is worth mentioning that several additional factors in our dataset also carried independent contributions to the predictability of mortality such as blood tests and imaging, which is in line with other studies72–76. These factors proved high potential to be incorporated in multiple scoring systems to improve mortality prediction, decision making, and logistical planning during the pandemic34,46,77. However, the time, effort, resources they require rendered them unavailable for a large number of patients in Syria. The added value they also brought to the scoring system did not justify dramatically reducing the sample size to include them. Despite that, six of the scores that incorporated blood tests could be obtained in a subgroup of our sample and still performed significantly poorer than LR-COMPAK. This reflects a dependency of the performance of mortality scores on the settings, where the sample is collected, highlighting the need for establishing and customizing scores based on locally available predictors in limited resources.
The potential of LR-COMPAK in forecasting mortality presents it as an objective valuable tool to assess cases’ severity and to rationalize and fasten hospitalization decisions especially in setting with limited inpatient capacity and scarce supplies 78. Its predictability performance was robust for all the included cities and medical centers without being limited to patients with positive PCR test27. Additionally, its performance extended to patients managed at home due to the limited hospitalization capacity with excellent sensitivity and specificity for forecasting mortality. This versatility underscores LR-COMPAK’s additional potential to enhance the coordination between different healthcare facilities, a challenge that represented a major obstacle towards enhancing hospitalization outcomes in COVID-19 patients27. LR-COMPAK can, for instance, allow for the comparison of case severity among hospitals and at home, guide the admission and referral of cases, and optimize the allocation of critical supplies. These potentials are inferred by several findings of our dataset. First, LR-COMPAK showed significant discrepancies between the patients presenting to the different hospitals. This explained, to some extent, the amplitude of mortality in centers like Al-Mowasat Hospital, which received the most severe patients on average and had the highest mortality rate. Second, the variance in mortality that could not be explained by the severity of the cases (i.e., high LR-COMPAK) can be due to the interaction between hospitals’ capacity, cases’ flow, and the available assets. For instance, patients in Aleppo had higher severity scores in comparison with Damascus and Lattakia, but they showed lower mortality rate. This observation may be attributed to the relatively large capacity of Aleppo University Hospital in light of the substantial decrease in Aleppo's population following the armed conflicts. Approximately one third of Aleppo's residents relocated to other cities or countries, a pattern that is not paralleled in the other studied cities79. Third, when plotting the mortality rate and LR-COMPAK across the study timeline, it was obvious that mortality rates spiked at both pandemic peaks in line with tendencies worldwide80–82. However, this pattern was not paralleled in the mortality score. This mismatch is most likely due to the strained resources by the higher flow of cases during these peaks, which can be alleviated by flexible resources and cases allocation among centers, especially after implementing a generalized scoring system of cases’ severity like ours.
Although patients’ flow was beyond hospitalization capacity, the mismatch was even wider between patients with an indication for intensive care and ICU beds in Syria26,27 and worldwide28. This constrained capacity forced the treating physicians in every center to subjectively triage cases selecting those more probable to benefit for ICU admission. However, the available guidelines developed to standardize this approach offered theoretical guidance rather than practical construction83. Also, LR-COMPAK did not exhibit an adequate predictability for mortality during ICU admission. Therefore, we developed a modified scoring system (LR-ALBO-ICU), which still included age and peripheral blood oxygen saturation but in combination with LDH and HCO3, two blood tests, which were repeatedly reported to reflect COVID-19 infection severity84–87. This score could satisfactorily isolate patients with high probability of survival in the ICU. Its accuracy of predicting ICU mortality (AUC of 0.77) was higher than three out of the four scores applied for the same aim (RECOILS (AUC of 0.75), 4C (AUC of 0.79), Sofa (AUC of 0.66), and SAPS-III (AUC of 0.72))46,88 with less demanding components than all the other scores46. These findings suggest that predicting mortality in resource-limited settings may be more straightforward and effective when compared to environments with ample resources. This difference may be attributed to the limited survival-protective impact of healthcare services, which may pale in comparison to the severity of cases upon presentation and the constrained care provided27. However, further studies are warranted to validate these findings in countries with similar sparse resources.
Limitations:
This study was not free of limitations. First, not all the patients included in our dataset had PCR positive tests due to the limited number of PCR kits89. However, we have shown previously a similar pattern of presentation and evolution of cases with and without PCR tests in Syria27. Also, all the reported analyses were replicated for both groups of patients to ensure the translatability of the developed scores. Second, our analysis could not reflect mild to moderate cases, which were mostly denied admission at presentation. However, LR-COMPAK showed excellent performance on another set of patients who were treated at home due to the limited hospitalization capacity90. Third, LR-ALBO-ICU relied on variables measured upon first presentation, which might not reflect the status of the cases when transferred to the ICU. However, the median delay from presentation to ICU admission was 0 (0,2), reducing the probability of a major change of the used evaluations. Forth, It has been internationally recommended that the triage procedure, which aims to prioritize patients in a way that maximizes benefits for the largest possible number of patients91 should be carried out by either a dedicated triage committee or, at the very least, a senior intensivist91. Nevertheless, the limited human resources in our setting prevented this and such decision had to be sometimes taken by second- to third-year internal medicine residents92. Despite that, we believe that the large sample size we reported from multiple centers all over the country constrained the effect of these limitations and provided a comprehensive dataset that allowed the development of robust scores that proved applicable in all the included centers.