In the frenzied early days of the COVID-19 pandemic, clinicians grappled with a perplexing puzzle. Patients who arrived with seemingly mild respiratory symptoms suddenly became critically ill, overwhelming intensive care units. A new body of research using machine learning promises to solve this mystery by predicting ICU admission from subtle early signals. While a specific 2026 Frontiers in Medicine study on machine learning for patients with mild respiratory failure remains behind access hurdles, parallel studies reveal consistent patterns across datasets, highlighting the algorithm’s potential to transform triage.
Studies have consistently identified age, comorbidities such as cardiovascular disease, and laboratory markers such as elevated lactate dehydrogenase (LDH), C-reactive protein (CRP), and D-dimer as key predictive factors. For example, Frontiers in Digital Health analyzed 66 parameters of coronavirus patients admitted to ICU and identified 15 key indicators, including gender, age, blood urea nitrogen (BUN), creatinine (Cr), INR, albumin, and history of neurological, respiratory, and cardiovascular disease, which achieved high predictive power on admission.
Core forecasters emerge from the flood of data
Random forests and gradient boosting machines dominate these models and outperform traditional scores. The 2025 Frontiers in Artificial Intelligence study trained on 10,378 Emory patients predicted ventilator, ECMO, and mortality rates, highlighting the need for early data-driven escalation. “Researchers have sought to develop data-driven mechanisms to predict the outcome of COVID-19,” the authors noted, citing previous studies on MV duration and ICU stay.
The XGBoost model has performed well in recent validation. Biomedicines report achieved 87% sensitivity, 85% specificity, and 0.95 AUC for ICU prediction using demographics, NLR, PLR, and CRP. Using ICD-10 codes in a cross-sectional study in Health Science Reports, Iranian researchers found that Naïve Bayes and LightGBM were superior because symptoms such as respiratory distress and comorbidities increase the need for ICU, stating that “timely identification of patients who require admission to the intensive care unit could potentially save lives.”
Algorithms that surpass clinician intuition
We emphasize that performance metrics are superior to heuristics. A Frontiers in Public Health meta-analysis of AI for severe prognosis of coronavirus showed that deep learning models from high-income settings produced higher specificity in ICUs. Models like PLOS One used deep neural networks on 5,766 patients to link procalcitonin, LDH, CRP, and ferritin to ICU and mortality risk. “Elevated ferritin is associated with acute respiratory distress syndrome.”
External validation remains important. The JMIR study of 12,000 patients used a random forest across variants and vaccination to achieve an AUC of 0.95 for mortality and perfect sensitivity in the ICU. However, as noted in the Iranian ML study, limitations still exist, such as data imbalance, missing values such as LDH in early records, and cohort specificity.
Real-world implementation hurdles
Interpretability is required for integration into workflows. SHAP analysis of recent papers revealed the importance of features consistent with biology (dyspnea, oxygen requirements, leukocytosis). The 2025 Frontiers in Public Health model defined critical illness as ventilation, ICU, or death and flagged elevated LDH and CK-MB. According to the study, “LDH has been reported to be elevated in critically ill patients treated in the ICU.”
Beyond COVID-19, these tools are evolving. The contribution to X focuses on ML for mortality and ARDS prediction after acute pneumonia and suggests broader respiratory applications. Frontiers in Medicine’s deep learning effort uses CT and laboratory to predict ARDS after ICU admission, driving multimodal fusion.
The road to bedside precision
For mild cases, the focus of the Frontiers summary, the model could avoid surprises by scoring risks at the time of presentation. Collective evidence suggests that an AUC greater than 0.90 is achievable when routine vitals, laboratory tests, and medical history are considered. As the IEEE study claims, ML enables “timely risk scoring and specific resource allocation.” Future iterations that combine EHRs and wearables could make ICU surprises a relic of the past.
