Evaluating the MPM III and SAPS III prognostic models in a war-affected, resource-limited setting: a prospective study from the Gaza Strip

Abstract

Background

Validation studies of prognostic models used in critical care have yet to be conducted in Palestine. The intense conflict in the Gaza Strip presents an opportunity to evaluate the performance of local ICUs and validate the performance of the MPM and SAPS models within a resource-limited and highly stressed healthcare system.

Methods

A prospective study conducted from October to December 2024 included all patients admitted to ICUs in three of the four critical care units operating in the Gaza Strip. Sociodemographic, clinical, physiological, and laboratory parameters were collected, along with information regarding the clinical course and ICU outcomes. The MPM-III and SAPS-III scores were calculated, and their discrimination and calibration were assessed using AUROC and the Hosmer-Lemeshow test, respectively. Furthermore, the difference between the predicted and actual mortality rates was visualized, and standardized mortality rates (SMR) were calculated. Except for the Hosmer-Lemeshow test, a p-value of less than 0.05 was deemed statistically significant. All statistical analyses were conducted using R Studio.

Results

The cohort included 101 patients, of whom 72.27% were surgical cases and 58.41% were admitted from the ER. The ICU mortality rate was 30.69%. The median duration of ICU admission was four days [IQR 2–9] and was significantly longer for surgical cases than for medical cases. Physiological and laboratory parameters, along with interventions associated with higher mortality, included a lower GCS, burns, elevated leukocyte and platelet counts, lower PPO2, dysrhythmias, intracranial mass effect, and the need for mechanical ventilation or central venous catheterization. The predicted mortality rates were 16.63% for MPM0-III and 16.82% for SAPS-III. SMRs indicated that both models underestimated ICU mortality (SMR, MPM0-III 1.85; SAPS-III 1.83), with the discrepancy more likely to occur in high-risk patients. ROC curves demonstrated acceptable to good discriminatory power for both models (AUROC, MPM0-III 0.79 (95% CI 0.7–0.88); SAPS-III 0.87 (95% CI 0.80–0.94)). The Hosmer-Lemeshow test yielded statistically insignificant results for both models, indicating good calibration.

Conclusion

The outcomes of critical care units in the Gaza Strip during the studied period of the war were comparable to those of other hospitals in the West Bank and other LMICs without active conflicts. The MPM-III and SAPS-III demonstrated good discrimination and calibration, making them valid tools for enhancing ICU performance and improving resource utilization in the Gaza Strip.

Introduction

Several predictive models have been developed and utilized in critical care settings for prognostication, benchmarking, and research standardization [1]. These models draw on data from the patient’s history, as well as physiological and laboratory parameters, which are input into a formula to calculate the most likely outcomes. The Acute Physiology and Chronic Health Evaluation (APACHE), the Simplified Acute Physiology Score (SAPS), and the Mortality Probability Models (MPM) along with their iterations represent some of the most commonly used predictive models due to their generally good ability to distinguish between high-risk and low-risk patients (discrimination) and to provide risk estimates that are not far from the actual mortality (calibration) [1,2,3]. However, these models have significant limitations despite being validated in various settings and health systems. For instance, they have primarily been developed in high-income (HIC) and resource-rich countries. Additionally, they variably rely on laboratory parameters that may not be readily available in low-income countries (LICs) or during conflicts or disasters. Furthermore, their calibration can deteriorate over time due to changes in case mix, which necessitates updates to keep pace with contemporary practice [1, 4].

ICU mortality in HICs ranges from 8 to 18%, varying between countries and hospitals, with higher rates for certain critical illnesses compared to the overall critically ill population [5, 6]. However, it is significantly higher in low- and middle-income countries (LMICs), reaching up to 53% in some reports [7]. This disparity is attributed to a combination of factors related to resources, healthcare providers, healthcare systems, and the overall critical care bed capacity in LMICs [5, 6]. Additionally, these factors, along with constraints on patient selection and infection control associated with conflict and disaster situations, further contribute to the increased ICU mortality rates in those settings.

In Palestine, a single-center study from the West Bank reported an ICU mortality rate of 31.7% [8]. However, those figures are not directly applicable to ICUs in the Gaza Strip, which has suffered from worse socioeconomic and healthcare conditions compared to the West Bank, even before the Gaza War that erupted in 2023 [9, 10]. The Gaza Strip is a chronic conflict zone that has been blockaded by Israel for nearly two decades, leaving its healthcare sector on the verge of collapse [11]. Between October 2023 and January 2025, it experienced a protracted, violent, and intense war that resulted in tens of thousands of Palestinians killed, over a hundred thousand injured, and the vast majority of the population displaced, hungry, and severely traumatized [12,13,14,15]. The war has pushed the Gaza healthcare sector to its limits, with a continuous influx of traumatic injuries, targeted assaults on healthcare facilities and staff, a severe shortage of resources, and ongoing pressures on the healthcare workforce, leaving it in a state of virtual collapse [16,17,18,19,20,21,22].

Although the performance of ICU scoring models has been extensively validated, most validations were conducted in HICs, and no previous assessments have been performed in Palestine or the Gaza Strip [23]. Furthermore, the literature on the performance of these models in ICUs operating during active conflicts is limited. The aims of the present study are twofold: first, it seeks to outline the patterns and outcomes of ICU admissions in the Gaza Strip during the war, and second, to validate the performance of two ICU scoring systems in the unique low-resource and conflict settings presented in the Gaza Strip.

Methods

Design and settings

This prospective study included patients admitted to the ICU at three tertiary hospitals in the Gaza Strip from October 15 to December 4, 2024. Exclusion criteria included patients with fixed and dilated pupils, those with significant missing data, and individuals who did not consent to participate.

At the time of this study, the Gaza Strip was divided by the Israeli military into completely isolated northern and southern enclaves, with mass displacement of the population to the southern enclave at the war’s onset. The northern enclave comprised the North and Gaza Governorates, with an estimated population of half a million out of an original 1.2 million residents before the war [24]. The southern enclave included the Middle, Khan Younis, and Rafah Governorates, with an estimated population of one and a half million during the war. Additionally, at the time of this study, the North and Rafah Governorates were experiencing significant Israeli ground offensives, which resulted in the suspension of health services and the displacement of residents to the remaining governorates within each enclave. The study settings comprised the ICUs of Al-Ahli Baptist Hospital, Al-Aqsa Hospital, and Nasser Medical Complex. The ICU at Al-Ahli Baptist Hospital had three beds and was the only critical care unit operating in a government-run hospital in the northern enclave at that time. The ICU at Al-Aqsa Hospital had a capacity of fourteen beds, serving as the only operational ICU in the Middle Governorate, which housed nearly three-quarters of a million residents and IDPs at that time. The ICU at Nasser Hospital had ten beds and was one of two ICUs functioning in the Khan Younis Governorate, the other being located at the European Gaza Hospital, which had sixteen beds. The three ICUs included in the study maintained an occupancy rate of 100% throughout the data collection period. All three ICUs were mixed closed unit ICUs managed by an intensivist and staffed by intensivists and medical doctors. None employed specific ICU admission criteria or severity scores in daily practice.

Instruments and procedures

A data collection sheet comprised of five sections was created for this study. The first section collected sociodemographic information such as age, sex, admission route and date, case type, and comorbidities.

The second section included admission variables according to the Mortality Probability Model (MPM0-III) severity score, which is calculated from age and 15 other dichotomous parameters measured on or within 1 h of ICU admission to provide the predicted mortality [2]. The MPM0-III has good discrimination and calibration and has been externally validated in additional ICU populations [1, 25, 26]. It also has the benefit of relying on clinical and physiological data only. The third section included the Simplified Acute Physiologic Score (SAPS-III), which is calculated from 20 parameters, including admission-related details, patient characteristics, and physiological and laboratory variables [3]. It has been externally validated in different populations [1, 27]. MPM0-III and SAPS-III are easier to extract and less reliant on laboratory tests than other models, such as the APACHE score. They are thus less labor-intensive and less susceptible to the limitations in laboratory investigations commonly encountered in conflict zones. Bilirubin was missing in all three ICUs and was coded as the reference “normal” category [1, 27].

The fourth section gathered additional information related to the cause of admission, including the site(s) of injury in surgical cases and the affected organs/systems in medical cases. The fifth section assessed the admission clinical course, encompassing infection, culture results and sensitivity profiles, antibiotic use, interventions (such as mechanical ventilation, inotrope use, laparotomy, thoracotomy, limb amputations, chest tube and central line insertion, splenectomy, casting, and open reduction and internal fixation “ORIF”), venous thromboembolism (VTE) prophylaxis, stress ulcer prophylaxis, analgesic administration, cardiopulmonary resuscitation (CPR), length of ICU stay (LoS), and the outcome at discharge from the ICU (dead/alive). Community-acquired infections (CAI) were defined as those manifesting within 48 h of admission, and hospital-acquired (HAI) infections if manifesting after 48 h [28].

Data collection was prospectively conducted by senior medical students from medical files and ended when the last recruited patient was discharged from the ICU. The study methods and procedures were piloted on ten patients, and after few modifications were made, the pilot results were excluded from the final analysis.

Ethical considerations

Ethical approval for the study was granted by the Institutional Review Board (IRB) at the Islamic University of Gaza. The research was conducted per the Declaration of Helsinki. Administrative approval was obtained from the Palestinian Ministry of Health. Informed written consent was given by patients or, in unconscious or confused patients, by the next of kin. No identifying information was collected, and anonymity was guaranteed.

Statistical analysis

The primary outcome of interest was mortality at the end of the ICU stay. Numerical data were reported as mean ± standard deviation (SD) or median and interquartile range (IQR), as appropriate. Categorical data were reported using frequency and percentage. Univariate analysis was conducted using the Mann-Whitney U test, Student’s t-test, chi-square test, and Fisher’s exact test, as appropriate.

Both study models were validated by assessing their discrimination and calibration. Discrimination refers to the model’s ability to differentiate between patients who experience the outcome and those who do not. This was evaluated using the area under the receiver operating characteristic curve (AUROC). Calibration, which measures the agreement between observed and predicted outcomes, was assessed using the Hosmer-Lemeshow test. A p-value greater than 0.05 in the Hosmer-Lemeshow test suggests acceptable calibration. Additionally, we visualized the difference between the predicted mortality rate, divided into quartiles (x-axis), and the observed mortality rate (y-axis). Standardized mortality rates (SMR) were calculated by dividing the observed death rate by the expected death rate and were reported with their corresponding 95% confidence intervals. A p-value of less than 0.05 was considered statistically significant. All statistical analyses were performed using R Studio.

Results

The final analysis included all 101 patients admitted to three hospitals during the study period. The median age of the cohort was 26 years (range 19–35), with 34.65% being female. Most patients were surgical cases (72.27%) and were admitted from the emergency room (58.41%). An ICU mortality rate of 30.69% was observed, with a median LoS of 4 days [IQR 2–9]. Surgical cases had longer ICU admissions compared to medical cases (median LoS: surgical 6 days [IQR 3–10]; medical 2 days [IQR 2–4], p < 0.001).

The admission Glasgow Coma Scale (GCS) was strongly associated with mortality (median GCS scores: survivors 14 [IQR 9–15]; non-survivors 7 [IQR 5–10], p < 0.001). Burn victims were also significantly more likely to die (p = 0.036). Table 1 summarizes the baseline characteristics of the study sample.

Table 1 Baseline characteristics of the study cohort

The univariate analysis indicated that surviving patients had lower leukocyte and platelet counts and higher partial oxygen pressure than non-survivors. Table 2 presents the physiological and laboratory parameters of the study cohort, conducted immediately upon or within one hour of admission to the ICU unit.

Table 2 Vital signs and laboratory values of the study cohort

Table 3 outlines the complications and interventions experienced by patients in the ICU. Twenty patients (19.80%) experienced infections, with 12 community-acquired and eight nosocomial infections. Antibiotics were administered to 98 patients (97.02%), while only nine cases were treated with antibiotic therapy guided by microbial cultures. Cardiac dysrhythmia and intracranial mass effect were detected in nine (8.91%) and 30 (29.70%) patients, respectively, and both were associated with higher mortality (p = 0.003). Invasive mechanical ventilation (MV) was performed on 72 (71.28%), and a central venous catheter (CVC) was placed in 55 patients (54.45%). Both were linked to higher mortality (p = 0.001 and p = 0.015, respectively). Sixty-seven patients (66.33%) were given VTE prophylaxis, while 96 (95.04%) were given stress ulcer prophylaxis.

Table 3 Acute events and interventions

Table 4 summarizes the performance of both mortality prediction models. The predicted mortality rates were 16.63% (SD 20.39) for MPM0-III and 16.82% (SD 15.52) for SAPS-III. The Standardized Mortality Ratio (SMR) indicated that both models underestimated ICU mortality (SMR, MPM0-III 1.85 (95% CI 1.25–2.62); SAPS-III 1.83 (95% CI 1.24–2.59)). Figures 1 and 2 illustrate the differences between the observed and predicted mortality rates distributed in quartiles for MPM0-III and SAPS-III, respectively. The discrepancy between observed and predicted mortality was more likely to occur in high-risk patients (Quartiles 3 and 4). The ROC curve (Fig. 3) demonstrated acceptable and good discriminatory power for MPM0-III and SAPS-III, respectively (AUROC, MPM0-III 0.79 (95% CI 0.7–0.88); SAPS-III 0.87 (95% CI 0.80–0.94)). The Hosmer-Lemeshow test was statistically insignificant for both models (MPM0-III: chi-square statistic = 5.762, p = 0.676; SAPS-III: chi-square statistic = 13.281, p = 0.102).

Table 4 Performance of mortality prediction models

Fig. 1

Observed versus expected mortality rate by MPM0-III prediction quartiles

Fig. 2

Observed versus expected mortality rate by SAPS-III prediction quartiles

Fig. 3

Receiver operating characteristic curve for both prediction models

Discussion

To our knowledge, this is the first study to report ICU admission patterns and outcomes during a highly intense and violent active war. It is also the first to explore the validity of ICU prognostic models, namely the MPM0-III and SAPS-III, for use in the Gaza Strip’s low-income and resource-depleted healthcare system. The study found an ICU mortality rate of 30.69%. Surgical cases accounted for the majority of admissions and had significantly longer ICU stays. Several physiological and laboratory parameters and interventions were associated with higher mortality, including a lower GCS, burns, elevated leukocyte and platelet counts, lower PPO2, dysrhythmias, intracranial mass effect, and the requirement for MV or a CVC. SMR analysis revealed that both models underestimated ICU mortality, while ROC curves demonstrated acceptable and good discriminatory power for MPM0-III and SAPS-III, respectively.

Before the current war, critical care in Gaza Strip hospitals faced an unprecedented set of challenges due to decades of blockade, movement limitations, and repeated Israeli military operations that have devastated the infrastructure, leaving healthcare vulnerable and chronically inadequate [29, 30]. The war that began in October 2023 further complicated this fragile situation by introducing new man-made obstacles that contributed to the collapse of healthcare services. For example, more than 6% of the Gaza Strip’s population has been killed or wounded so far in the war, overwhelming health services and critical care units [31, 32]. ICUs were also affected by a severe shortage of sanitation materials, hindering infection control [32]. Additionally, geospatial analysis has shown a staggering number and scale of direct attacks on or near healthcare facilities in the war’s initial two months, with several attacks directly targeting ICUs in various hospitals [16, 20, 32]. Healthcare personnel have also experienced direct assaults, with over a thousand healthcare workers killed, many while actively serving in hospitals [31, 33]. Furthermore, movement restrictions and limitations on the entry of medical supplies and medications have severely hindered healthcare delivery, including critical care provision [32]. Moreover, a significant number of skilled professionals have emigrated in the early months of the conflict, while those who remain are facing high levels of burnout and financial strain [18, 22, 34].

People in areas experiencing armed conflict continue to suffer from everyday medical complications. However, the stress on healthcare systems and the rise in surgical cases due to violence consume a significant portion of the available critical care capacity. This is sharply reflected in the present study, where emergency surgical cases comprised 66.33% of admissions, and surgical cases overall accounted for 72.27%. At the same time, the remainder were primarily medical cases, with only one obstetric admission involving a patient with eclampsia. In contrast, a recent report from ICU units in seven conflict zones found that trauma cases accounted for a mere 18% of admissions. On the other hand, obstetric, pediatric, and medical cases represented 34%, 26%, and 8%, respectively [35]. Given the absence of dedicated obstetric or pediatric ICUs during the present study’s period, the reported admission patterns likely reflect the volume of trauma cases in hospitals. A possible tendency for local crews to admit surgical cases or those who were more likely to survive could also have played a role, but this is impossible to verify since no specific ICU admission criteria were available. Additionally, based on the authors’ own experience and due to nearly continuous 100% ICU occupancy, managing critical cases in hospital wards or in emergency and recovery rooms became a common measure [34].

The ICU mortality rate of 30.69% in the present study is close to the 31.7% reported in the West Bank by Tambour et al. [8]. Similar rates were also reported from other LICs such as Kenya (31.7%) [7], whereas rates of 24% and 46.8% were reported from Egypt and Ethiopia, respectively [36, 37]. Nonetheless, those rates remain higher than the 16.1% average rate revealed in a large study involving 84 countries that assessed the global burden of critical illness [6]. Factors related to the specific case mix in the cohort likely influenced and reduced the current cohort’s mortality rates. For instance, in the aforementioned study on conflict areas, trauma cases had a lower mortality rate than obstetric, pediatric, and medical cases, which are underrepresented in the cohort [35].

Another possible factor is the relatively low infection rate of 19.80% observed in our cohort, which could be related to the younger age, unique case mix, shorter LoS, and the nearly universal administration of antibiotics upon admission. For instance, ICU mortality in patients in LMICs in the Middle East was 17.12% in the absence of HAIs, and 30.15% and 63.44% in patients with one and three HAIs, respectively [38]. Similarly, mortality rates of 39.7% and 54.5% were reported in septic ICU patients in a two-year, single-center, retrospective study from the northern West Bank and in another comparable study from neighboring Egypt [39, 40]. These studies also reported average LoS of eight and 12 days, respectively, compared to four in our study (two for medical cases), which reflects the differences in the case mix and the distinct conditions created by the war in Gaza. Lastly, the present study took place 12 months after the war’s onset, while the initial months were the most violent and had the greatest impact on healthcare services. Therefore, the results may not be directly generalizable to the earlier phases of the war.

Accurate discrimination and calibration are two critical characteristics that predictive scoring systems must achieve when used in critical care. The present study found acceptable and good discriminatory power for MPM0-III and SAPS-III, with AUROC values of 0.79 and 0.87, respectively. A study from Tanzania, an African LIC, reported AUROC values of 0.90 and 0.89 for MPM-III and SAPS-III, respectively, alongside an in-hospital mortality rate of 16.1%. In contrast, a study from Kenya reported an AUROC of 0.74 for MPM-III and an ICU mortality rate of 31.7% [7, 25]. A meta-analysis evaluating the performance of ICU scoring systems in LMICs found that excellent discrimination was reported more frequently when hospital mortality was the outcome, compared to ICU mortality [23]. This same study indicated an AUROC range of 0.71–0.94 for MPM-III and 0.80–0.94 for SAPS-III, with the frequency of reported “good” or “very good” discrimination being lower for MPM compared to SAPS, which aligns with our findings [23].

Regarding calibration, the present study found that both models were well-calibrated and did not overestimate mortality (p = 0.676 and p = 0.102). Both Tanzania and Kenya also reported good instrument calibration [7, 25]. In the aforementioned meta-analysis, good calibration was reported in 66.7% of SAPS-III model performance evaluations, whereas for MPM-III, only two out of four studies evaluating MPM-III reported p-values above 0.05 [23]. However, good calibration was noted more frequently when ICU mortality was the outcome than when hospital mortality was used [23]. Overall, good discrimination and good calibration were reported together in only 11.9% of the same evaluations.

The Hosmer–Lemeshow goodness-of-fit test results should be interpreted with caution in small cohorts, as the test may yield insignificant results due to low statistical power [41]. We calculated the SMR and visualized the discrepancy between observed and expected mortality across quartiles to enhance our understanding of calibration in the present cohort. Both models underestimated mortality, with SMR values of 1.85 and 1.83, particularly among high-risk patients. A study from Brazil reported SMR values of 1.29 for SAPS-III and 3.41 for MPM-III, along with an in-hospital mortality rate of 30.4% [42]. Another study indicated that both models tend to overestimate in-hospital mortality, with SMR values of 0.61 for MPM-III and 0.46 for SAPS-III, in a cohort with a relatively low mortality rate of 9.1% [43].

The findings of this study suggest that the model incorporating physiological and laboratory parameters (SAPS-III) demonstrates better performance in the ICUs of the Gaza Strip compared to MPM-III, which relies solely on clinical and physiological data. However, wartime conditions and low-resource settings may limit the applicability of the former due to challenges in laboratory testing, making MPM-III a feasible alternative. Healthcare authorities and stakeholders are encouraged to integrate these findings into future policies and ICU quality improvement strategies to optimize resource allocation and inform treatment efforts, which are crucial in chronic conflict areas such as the Gaza Strip.

Strengths and limitations

The current study collected data from and validated two ICU prognostication models in hospitals operating during a unique humanitarian crisis. This is also the first such study conducted in the Gaza Strip. The prospective design minimized data loss and facilitated meaningful results. However, it has several limitations. Due to logistical constraints, the sample size was relatively small, and one ICU unit could not be included. Additionally, more sophisticated models that relied more heavily on laboratory parameters (e.g., APACHE score) were not evaluated due to the same limitations. Furthermore, while the study reported ICU outcomes, including hospital outcomes might provide deeper insights, as discharge from the ICU could be influenced by non-clinical factors such as a shortage of ICU beds. Lastly, the results on ICU admissions and outcomes may not accurately reflect the most violent periods, particularly during the first few months of the war.

Conclusions

The outcomes of critical care units in the Gaza Strip during the studied period of the war appear to be comparable to those of other hospitals in the West Bank and other LMICs without active conflicts, despite a predominance of traumatic cases and the depletion of healthcare services. However, other periods of the war might have experienced different circumstances and case mixes, possibly leading to less favorable outcomes. The study invites strengthening data infrastructure in the Gaza Strip’s critical care units in order to facilitate benchmarking and research. Moreover, incorporating clear ICU admission criteria and context-sensitive prognostic models into daily practice can enhance ICU performance in the Gaza Strip and improve resource utilization. Both the MPM-III and SAPS-III can serve as valid tools that demonstrate good discrimination and calibration in the Gaza Strip’s unique circumstances. Furthermore, to gain a more comprehensive understanding of critical care performance, future studies should incorporate post-ICU hospital outcomes, which will help pinpoint systemic barriers to recovery and inform patient-centered care.