Building an ICU Early Warning Pipeline with Clinical Scores

Early warning scores detect clinical deterioration hours before a cardiac arrest or a sepsis crisis. NEWS2 was adopted across the UK NHS specifically because it reduces unexpected ICU admissions, code blue events, and mortality. qSOFA screens for sepsis using three bedside observations that require zero lab work. SOFA quantifies organ failure severity on a 0 to 24 scale.

Most hospitals track vitals. Almost none compute these scores automatically. The data is there. The scoring is not.

Here is a pipeline that takes a table of ICU patient vitals, computes all three scores for every patient, and generates a priority sorted risk report.

The Pipeline

pip install moisscode

from moisscode.modules.med_scores import ClinicalScores


def score_patients(patients):
    """
    Score a list of patient vital sign dicts.
    Each patient dict needs: id, rr, sbp, hr, temp, spo2, gcs, on_o2, avpu
    Returns scored patients sorted by risk (highest first).
    """
    results = []

    for p in patients:
        patient_obj = type("Patient", (), p)()

        qsofa = ClinicalScores.qsofa(patient_obj)
        news2 = ClinicalScores.news2(patient_obj)

        risk_level = news2["risk"]
        action = news2["action"]

        results.append({
            "id": p["id"],
            "qSOFA": qsofa,
            "qSOFA_positive": qsofa >= 2,
            "NEWS2": news2["score"],
            "NEWS2_risk": risk_level,
            "action": action,
            "vitals": {
                "RR": p["rr"], "SBP": p["sbp"], "HR": p["hr"],
                "Temp": p["temp"], "SpO2": p["spo2"], "GCS": p["gcs"],
            },
        })

    # Sort by NEWS2 score descending (sickest first)
    results.sort(key=lambda x: x["NEWS2"], reverse=True)
    return results


def print_report(scored):
    """Print a clinical risk stratification report."""
    print(f"{'ID':<8} {'qSOFA':<8} {'NEWS2':<8} {'Risk':<10} {'Action'}")
    print("-" * 70)
    for p in scored:
        flag = "⚠" if p["qSOFA_positive"] else " "
        print(
            f"{p['id']:<8} {p['qSOFA']}{flag:<7} {p['NEWS2']:<8} "
            f"{p['NEWS2_risk']:<10} {p['action']}"
        )

Running It

Five ICU patients. Vitals pulled from bedside monitors:

patients = [
    {"id": "P001", "rr": 18, "sbp": 120, "hr": 78,  "temp": 37.0,
     "spo2": 97, "gcs": 15, "on_o2": False, "avpu": "A"},
    {"id": "P002", "rr": 26, "sbp": 88,  "hr": 112, "temp": 38.8,
     "spo2": 91, "gcs": 14, "on_o2": True,  "avpu": "V"},
    {"id": "P003", "rr": 32, "sbp": 82,  "hr": 130, "temp": 39.2,
     "spo2": 88, "gcs": 11, "on_o2": True,  "avpu": "V"},
    {"id": "P004", "rr": 14, "sbp": 135, "hr": 68,  "temp": 36.8,
     "spo2": 98, "gcs": 15, "on_o2": False, "avpu": "A"},
    {"id": "P005", "rr": 22, "sbp": 100, "hr": 95,  "temp": 38.1,
     "spo2": 94, "gcs": 13, "on_o2": False, "avpu": "A"},
]

scored = score_patients(patients)
print_report(scored)

Output:

ID       qSOFA    NEWS2    Risk       Action
----------------------------------------------------------------------
P003     3⚠      13       HIGH       Emergency assessment by clinical team
P002     3⚠      12       HIGH       Emergency assessment by clinical team
P005     3⚠      8        HIGH       Emergency assessment by clinical team
P001     0       0        LOW        Continue routine monitoring
P004     0       0        LOW        Continue routine monitoring

Three patients need emergency assessment. P003 and P002 are the sickest (NEWS2 of 13 and 12). P005 is deteriorating (NEWS2 of 8, all three qSOFA criteria met). P001 and P004 are stable.

The report is sorted by severity. The sickest patients are at the top. In an ICU with 30 patients and 2 nurses, this sorting determines who gets attention first.

Plugging Into a DataFrame

For researchers working with patient cohort data in pandas:

import pandas as pd

df = pd.DataFrame(patients)

df["qSOFA"] = df.apply(
    lambda row: ClinicalScores.qsofa(type("P", (), row.to_dict())()), axis=1
)

df["NEWS2"] = df.apply(
    lambda row: ClinicalScores.news2(type("P", (), row.to_dict())())["score"], axis=1
)

print(df[["id", "qSOFA", "NEWS2"]].sort_values("NEWS2", ascending=False))

     id  qSOFA  NEWS2
2  P003      3     13
1  P002      3     12
4  P005      3      8
0  P001      0      0
3  P004      0      0

Two new columns added to the DataFrame. Every sepsis prediction or ICU mortality model published in Indian Journal of Critical Care or IJMS reimplements these scores from scratch. This is a validated, tested implementation that runs across an entire dataset with df.apply.

Available Scores

The scoring engine has 13 validated systems. Any of them can be plugged into this pipeline:

Score	Domain	Inputs
qSOFA	Sepsis screening	RR, SBP, GCS
SOFA	Organ failure (0 to 24)	PaO2/FiO2, platelets, bilirubin, MAP, GCS, creatinine
NEWS2	Deterioration detection (0 to 20)	RR, SpO2, temp, SBP, HR, AVPU, O2 status
APACHE II	ICU mortality (0 to 71)	Vitals + labs + age + chronic health
CHA2DS2 VASc	Stroke risk in AF	Age, sex, CHF, HTN, stroke history, vascular disease, diabetes
HEART	Chest pain stratification	History, ECG, age, risk factors, troponin
MELD	Liver transplant priority	Bilirubin, creatinine, INR, sodium
Child Pugh	Cirrhosis severity	Bilirubin, albumin, INR, ascites, encephalopathy
CURB 65	Pneumonia severity	Confusion, urea, RR, BP, age
Wells PE	PE probability	Clinical signs, heart rate, immobilization, history
Glasgow Blatchford	Upper GI bleed	BUN, Hgb, SBP, HR, presentation
KDIGO AKI	Acute kidney injury staging	Baseline and current creatinine, urine output
Framingham	10 year CV risk	Age, sex, cholesterol, HDL, SBP, smoking, diabetes

Each one is a static method that takes a patient object and returns the score. No API keys, no internet, no subscription.

Professional biomedical software. All clinical outputs should be validated by qualified professionals before patient care application.

Scores Module Documentation | GitHub