The LACE Score: A Brief History
The LACE index was developed in 2010 by van Walraven and colleagues as a simple tool to predict 30-day readmission risk. It uses four variables: Length of stay, Acuity of admission, Comorbidities (Charlson index), and Emergency department visits in the past 6 months.
For its time, LACE was a reasonable approach. It required no specialized technology, could be calculated by hand, and offered a standardized way to identify high-risk patients. Many health systems adopted it as their default readmission risk tool.
The Limitations Become Clear
However, over the past decade, the limitations of LACE have become increasingly apparent:
1. Modest Discrimination
Multiple validation studies have shown LACE achieves a C-statistic (AUC) of approximately 0.60-0.68 for 30-day readmission prediction. This means the score performs only slightly better than chance at distinguishing patients who will be readmitted from those who won't.
To put this in perspective, an AUC of 0.50 is random chance—a coin flip. LACE's performance means that roughly 1 in 3 high-risk patients identified by LACE won't actually be readmitted, while many patients who will be readmitted are missed entirely.
2. Poor Calibration
LACE scores often don't translate to accurate absolute risk predictions. A patient with a LACE score of 10 might have a predicted risk of 25%, but the actual observed rate could be significantly different—sometimes by 10 percentage points or more.
This calibration gap matters because clinical decisions depend on understanding actual risk levels. If a care manager is told a patient has a 30% readmission risk, but the true risk is 18%, resources may be misallocated.
3. Missing Key Predictors
The LACE variables capture only a fraction of the information relevant to readmission risk:
4. One Size Doesn't Fit All
LACE was developed on a Canadian general medicine population. Its performance varies significantly across different patient populations, conditions, and healthcare settings. A score calibrated for medicine patients may overestimate or underestimate risk for surgical, cardiac, or oncology populations.
The Machine Learning Alternative
Modern machine learning approaches can address these limitations:
Incorporating more variables: ML models can process hundreds of structured data elements from the EHR, identifying complex patterns that simple additive scores miss. Marqi Index uses over 150 validated predictors.
Learning from local data: Models can be trained or fine-tuned on a health system's own patient population, improving calibration and relevance to the specific case mix.
Capturing non-linear relationships: The relationship between predictors and readmission isn't always linear. A 2-day length of stay might have different implications for hip replacement vs. heart failure. ML models can identify these nuances.
Continuous improvement: Unlike static scores, ML models can be updated as care patterns, patient populations, and clinical practices evolve.
What to Look For in an ML Alternative
Not all ML solutions are created equal. When evaluating alternatives to LACE, health systems should ask:
1. **Has the model been validated on external cohorts?** Internal validation alone isn't sufficient. Demand external validation on populations your organization didn't train on.
2. **Is the validation peer-reviewed?** Published validation studies provide accountability and transparency that marketing materials don't.
3. **What's the calibration performance?** AUC alone isn't enough—you need accurate absolute risk predictions for clinical decision-making.
4. **What data does it require?** Solutions that demand additional data capture create implementation barriers. Look for models that use data already in your EHR.
5. **How does it integrate?** Can it work with your existing EHR and workflows without requiring clinicians to use separate systems?
The Evidence for Marqi Index
Marqi Index was designed to address LACE's limitations directly:
Conclusion
LACE served an important role in bringing attention to readmission risk, but it's no longer the best option available. Health systems facing HRRP penalties and seeking to improve patient outcomes should evaluate validated ML alternatives that offer better discrimination, calibration, and clinical utility.
The evidence is clear: we can do better than LACE. The question is whether health systems are ready to make the transition to validated AI that actually performs at the level clinical decisions require.