The article discusses the potential of AI in flagging early sepsis cases for prompt antimicrobial care. It highlights the limitations of current models and the need for further validation before clinical adoption.

---

The prospect of a sepsis diagnosis instils fear in patients and clinicians alike. This deadly condition accounts for 11 million associated deaths among 49 million cases worldwide. The early detection of sepsis and appropriate workup and treatment are necessary to improve survival. However, presentations are highly heterogeneous, making it difficult to diagnose and treat this disease in a timely fashion. 

---

SOFA score

The Sequential Organ Failure Assessment (SOFA) score is a scoring system developed for sepsis research.[1][2] It gauges the function of the body’s organ systems, including neurologic, blood, circulatory, liver, and kidney, and assigns a score based on the data from each category.

A higher SOFA score is linked to higher mortality, with a score of  ≥2 suggesting life-threatening organ dysfunction secondary to infection. Such a score is correlated with an in-hospital mortality of 10%.

A bedside tool yielding the quick SOFA score (qSOFA) was developed for use in patients outside the ICU. As it relies on three criteria, this measure is referred to as the Sepsis-3 criteria. For a patient to be prone to a poor outcome secondary to sepsis, at least two of the three clinical variables need to be present: 

• Low blood pressure (SBP ≤ 100 mmHg)
• Increased respiratory rate (≥ 22 breaths per minute)
• Altered mentation (Glasgow Coma scale [GCS] score < 15) 

Catching sepsis early, although difficult, allows for easier management with antibiotics. If caught later, however, it’s much harder to treat—albeit easier to diagnose.

One of the many reasons behind the difficulty of early diagnosis, according to the authors of a narrative review in Computers in Biology and Medicine, is that “sepsis is a very heterogeneous syndrome. Patients may develop sepsis based on different pathophysiological mechanisms and may present with different clinical phenotypes. About one in five patients that present to the emergency department with suspected sepsis does not show any signs of organ dysfunction, while they will develop this within 48 h of admission.”

---

SOFA limitations

Extant screening tools for sepsis suffer from major limitations.[3] For example, qSOFA lacks sensitivity, whereas SOFA was modelled after populations and not patients.

“Though SOFA was developed for sepsis research and has been validated in additional settings, there is concern that it does not accurately predict mortality when used for patients with isolated respiratory failure as demonstrated during the 2009 H1N1 pandemic," explains Health & Human Services.

"Very few patients with primary respiratory failure generate SOFA scores over 4-6, severely limiting its utility in a pandemic/epidemic and potentially biasing against patients with other conditions. An additional limitation is that elevated baseline creatinine, and in particular pre-existing end-stage renal disease, can cause the score to be falsely elevated about actual mortality.”

---

The rationale for AI in sepsis

According to the authors of the narrative review, to enhance patient outcomes, it’s necessary to shorten the time to diagnosis and improve the accuracy of prognosis for patients with sepsis: “Some patients who are initially not even categorised as having sepsis might benefit greatly from early administration of antibiotics."

"AI prediction models, which have shown to be useful for diagnosing and prognostication in other fields of medicine, could potentially add much value to these areas for patients with sepsis."

— Authors, Computers in Biology and Medicine

---

AI models

AI models for the early diagnosis of sepsis vary in prognostic ability.

1. According to the authors of the narrative review, one algorithm predicted sepsis onset 48 hours in advance with an AUROC of 0.83, based only on vital signs.
2. Another test, titled the Risk of Sepsis score, attained an AUROC of 0.93 during the first hour of admission, which increased to 0.97 after 24 hours. 
3. Not all AI models have worked out, according to the authors of an article in the American Journal of Respiratory and Critical Care Medicine (AJRCCM).[4]
4. When researchers validated the Epic Sepsis Model (ESM), they found that it had poor discrimination, despite its implementation by hundreds of hospitals. The AUC was 0.63, with poor discrimination and calibration, which was much lower than initial reports of an AUC of 0.76–0.83. In other words, physicians needed to evaluate 109 patients to detect one case of sepsis earlier than if they did not use the ESM. 
5. A more promising sepsis-detection tool is the Targeted Real-time Early Warning System (TREWS) to improve patient outcomes, which has been in development for years. It eventually attained an AUC of 0.97 for detecting sepsis, catching 82% of cases.

---

Moving forward

Although AI promises earlier diagnosis of sepsis in patients and the prospect of direct life-saving treatment, successful wide-scale deployment of this technology remains a challenge for many reasons.

One reason, as the narrative review authors acknowledge, is that the generalisability of AI algorithms remains poor. Additionally, AI models can be biased due to predictor values figured into the outcome.

There is also insufficient data on which to build accurate models in clinical practice. Overall, there’s a lag between the creation and the clinical implementation of AI algorithms to predict sepsis.

The AJRCCM authors urge caution. The most concerning aspect of the tools for early sepsis detection, they say, is that “the evidence for their clinical benefits is still preliminary and circumstantial."

They note that high-quality trials to measure the impact of alerts on care processes are lacking, and that, while the TREWS score is the most reliable in evaluating/detecting sepsis, the "study design limits the conclusions that can be drawn.” 

"Implementing and adopting these tools in practice should not be taken lightly, because unjust use can cause harm."

— Authors, American Journal of Respiratory and Critical Care Medicine

They note that sepsis alerts "triggering one-size-fits-all protocols" can lead to the overuse of antibiotics, which may cause a "significant burden when physicians must evaluate many patients to detect one sepsis patient early."

"We need RCT-level evidence to unequivocally show that using these alerts will have a positive net benefit for the patient. Setting up such studies will present significant challenges, including the blinding, level of randomisation, regulatory barriers, and cohort selection,” they concluded.

---

Disclaimer: This story is contributed by Naveed Saleh and is a part of our Global Content Initiative, where we feature selected stories from our Global network which we believe would be most useful and informative to our doctor members.