What Is a Readmission Risk Score and How Does It Work?

Readmission risk scores are standardized tools that estimate how likely a patient is to return to the hospital within a set window after discharge, usually 30 days. The most widely used scores, such as HOSPITAL and LACE, typically achieve moderate accuracy, with discrimination values hovering around 0.65 to 0.73 on a scale where 1.0 would mean perfect prediction and 0.5 is no better than a coin flip. That middling performance has pushed researchers toward machine learning and richer data sources, but the gains have been surprisingly incremental, and the practical challenge of acting on a risk score remains at least as hard as computing one.

How the Most Common Scores Work

Two scores dominate hospital practice. The HOSPITAL score assigns points based on seven variables: hemoglobin level at discharge, whether the patient was discharged from an oncology service, sodium level at discharge, whether the procedure was coded as urgent, the number of hospital admissions in the prior year, length of stay, and the type of admission. An international validation across hospitals in the United States, Canada, Israel, and Switzerland found the HOSPITAL score performed best in Canada, with a c-statistic of 0.78, and somewhat lower in the U.S. at 0.72 and in Israel and Switzerland at 0.68 each.¹

The LACE index uses four inputs: length of stay, acuity of the admission, comorbidities measured by the Charlson index, and the number of emergency department visits in the prior six months. LACE is popular partly because it can be calculated quickly from administrative data, but its accuracy outside the population it was developed in has been disappointing. When six readmission scores were externally validated against a Swiss hospital cohort, LACE managed a c-statistic of just 0.53, essentially random, while the HOSPITAL score reached 0.73.² That gap matters because a score that works well at one institution can fail badly at another, and many hospitals have adopted LACE without verifying its local performance.

What These Scores Tend to Capture

Most readmission scores draw on some combination of demographics, prior healthcare use, chronic conditions, and lab values. Age, number of recent hospitalizations, length of stay, and comorbidity burden show up in nearly every model. Heart failure-specific tools add variables like ejection fraction, blood pressure at discharge, and whether certain medications were prescribed before the patient left.³ One heart failure scale called the RAHF divides patients into low, moderate, and high risk tiers, where the high-risk group had roughly 1.7 times the readmission rate of the low-risk group.⁴

These variables are useful on average but leave a lot of individual risk unexplained. A patient might have favorable lab values and no recent hospitalizations yet still be heading home to a chaotic living situation with no reliable transportation to follow-up appointments. That kind of context rarely appears in the structured fields a risk score can read.

Frailty as a Predictor

Frailty, the overall vulnerability that accumulates with aging, chronic disease, and functional decline, has a clear link to readmission, but it is hard to capture in routine data. In a Canadian study, moderate-to-severe frailty was independently associated with roughly double the odds of readmission or death within 30 days, even after accounting for age, sex, and LACE score.⁵ A Danish study of acutely admitted older patients found an 88 percent higher risk of readmission or death within six months for those with the highest frailty scores, though the prediction models that incorporated frailty still topped out at c-statistics around 0.72.⁶ A separate study using a multidimensional frailty index found that the most frail patients had about 2.6 times the readmission hazard of the least frail.⁷

The pattern is consistent: frailty is a strong risk marker, but adding it to a model lifts overall accuracy only modestly. This is a recurring theme across readmission prediction. Individual risk factors can matter a great deal for a given patient without dramatically improving a model’s ability to sort everyone correctly.

Do Machine Learning Models Do Better?

Researchers have thrown every modern algorithm at this problem, from gradient-boosted trees to deep neural networks, hoping the pattern-recognition power of machine learning would crack what simpler scores could not. The results are real but underwhelming. A meta-analysis comparing machine learning to standard logistic regression found that machine learning improved the area under the curve by an average of 0.03, with deep learning methods doing best at about 0.06 above logistic regression.⁸ In practical terms, a jump from 0.70 to 0.73 means a small number of additional patients are correctly sorted into risk categories, which can matter at scale but is not the breakthrough many expected.

A large validation study using electronic health records told a similar story. Gradient-boosted decision trees achieved a c-statistic of 0.764 when given access to over 1,500 variables, but standard logistic regression with regularization hit 0.755 on the same data, and the confidence intervals overlapped.⁹ One Chinese study using a back-propagation neural network did reach 0.74 with better sensitivity and specificity than LACE or HOSPITAL, but that was a single institution’s data and the results have not been widely replicated.¹⁰

The honest takeaway is that the ceiling on readmission prediction is not mainly a modeling problem. Much of what drives a readmission, such as whether a patient fills a prescription, whether they have someone at home who can help, or whether an unrelated illness strikes, simply is not in the health record at the time of prediction.

Unstructured Data and Clinical Notes

One promising direction for pushing past that ceiling is mining the free-text notes clinicians write. Discharge summaries, nursing assessments, and progress notes contain observations about a patient’s social situation, cognitive status, and treatment adherence that never make it into coded fields. Models trained on clinical notes using natural language processing have shown meaningful gains. A study using home healthcare nursing notes alongside traditional variables boosted the area under the curve from 0.739 to 0.824.¹¹ Hybrid models that combine structured electronic health record data with unstructured text have also outperformed text-only or structured-only approaches, achieving better precision and more balanced recall.¹²

The difficulty is operational. Extracting useful signals from millions of clinical notes requires heavy computational infrastructure and raises privacy concerns that structured data does not. Most hospitals are nowhere near deploying these systems at the bedside.

Patient-Reported Outcomes

Another data source that standard models ignore is what patients themselves report about their health. Patient-reported outcome measures, or PROMs, capture things like pain levels, mobility, anxiety, and ability to carry out daily activities. A systematic review concluded that PROMs have potential to improve readmission prediction, though study designs have varied widely enough that the evidence is not yet definitive.¹³ A recent study found that adding PROMs to administrative data raised the c-index from 0.762 to between 0.774 and 0.782, depending on which instrument was used, with the largest gains among patients with chronic conditions likely to lead to preventable readmissions.¹⁴

The appeal of PROMs is that they can surface issues invisible to clinical data, like a patient who looks medically stable but reports severe depression or inability to manage medications at home. The challenge is logistical: collecting reliable responses from patients during or after a hospital stay is harder than pulling data from an electronic record.

Social Determinants and Equity

Poverty, low education, unstable housing, and limited English proficiency all raise the risk of bouncing back to the hospital. You would expect adding these social determinants to a risk model to improve its accuracy, but the results have been mixed in ways that reveal something important about how these models work. One study found that patients with unfavorable social determinants had higher HOSPITAL scores, yet adding those factors to the model did not budge the overall c-statistic, which stayed at 0.71. Social determinants explained only about 0.2 percent of the variation in scores.¹⁵

The likely explanation is not that social factors do not matter, but that their effects are already partially captured by clinical proxies in the model: number of prior admissions, comorbidity burden, and emergency department use all correlate with social disadvantage. However, when researchers looked at specific subgroups, social determinants did improve prediction. Adding individual- and community-level factors improved accuracy for Medicaid patients, older adults, and patients with obesity, with c-statistics rising by 0.02 to 0.03 in each group.¹⁶ A study using the LACE index similarly found that adding social determinants improved its c-statistic from 0.698 to 0.708, with the largest gains among Black patients and those 65 and older.¹⁷

Algorithmic Bias in Risk Scores

Getting the overall accuracy number right does not mean a model treats all groups fairly. A study evaluating readmission models across two U.S. states found that Black and low-income patients had higher false positive rates, meaning they were more often flagged as high risk when they were not actually readmitted. White and higher-income patients had higher false negative rates, meaning they were more often missed when they did come back. Overall error rates were higher for Black and low-income groups.¹⁸

A separate analysis of multiple models found that the HOSPITAL score and a retrained version of the CMS readmission measure showed the lowest racial bias, while some other models amplified disparities. White patients again showed higher false negative rates and Black patients higher false positive rates across the board.¹⁹ The practical consequences cut both ways: flagging patients as high risk when they are not wastes scarce transition-care resources, while missing truly high-risk patients defeats the whole purpose of the score.

When to Run the Score

Most readmission scores are designed to run at discharge, when the most clinical data is available. But research suggests there is value in predicting risk at admission, too. Institution-specific models that used only pre-admission data already outperformed the LACE index, which waits for discharge-time variables. For all-cause readmission, an admission-time model achieved a c-statistic of 0.60 compared to 0.55 for LACE, and the gap widened for specific conditions.²⁰ Early identification gives hospital staff more time to arrange services like home health visits or medication counseling before the patient walks out the door.

A study directly comparing the two approaches found that the populations identified as high risk at admission and at discharge overlapped but were not identical, and recommended a two-time-point strategy: screen at admission with available data, then re-screen at discharge with the full picture.²¹ In pediatric settings, an at-admission model performed comparably to an at-discharge model, reinforcing the idea that early screening can be just as useful and allows more time for interventions targeting social determinants of children’s health.²²

How Risk Scores Guide Follow-Up Care

One of the main reasons to calculate a readmission risk score is to decide how aggressively to schedule follow-up. The evidence here is striking: early follow-up after discharge helps most for patients who are already at the highest risk. A study of outpatient follow-up timing found that a visit within 14 days cut readmissions by roughly 1.5 percentage points for the lowest-risk patients but by about 19 points for the highest-risk group. A follow-up within seven days was associated with meaningful reductions specifically for patients with multiple chronic conditions and a baseline readmission risk above 20 percent, a group comprising about a quarter of all discharged patients.²³

A similar pattern appeared in surgical patients: after high-risk thoracic aortic aneurysm repair, patients estimated to be at highest readmission risk had significantly lower readmission rates in regions with robust primary care follow-up, while follow-up timing made no significant difference for lower-risk surgical patients.²⁴ Not every study found a benefit, though. One analysis of over 9,600 discharges found that primary care follow-up within 14 days was not significantly associated with fewer 30-day readmissions after adjusting for other factors.²⁵ The key difference may be whether the follow-up visit is targeted at high-risk patients or applied broadly, which is exactly the kind of triage a risk score is meant to enable.

Financial and Policy Stakes

Behind the clinical rationale sits a financial one. Medicare’s Hospital Readmissions Reduction Program penalizes hospitals with excess readmission rates for conditions like heart failure, pneumonia, and hip and knee replacement. Those penalties create strong institutional incentives to identify and intervene with high-risk patients. Research on AI-based predictive models in Medicare hospitals has estimated annual savings of roughly $1.3 million to $2.3 million per hospital, alongside reductions in 30-day readmission rates of about 12 to 15 percent.²⁶

But the penalty structure itself has drawn criticism. A JAMA Internal Medicine analysis found that adjusting for additional patient characteristics beyond what Medicare uses, such as disability, mental illness, and socioeconomic disadvantage, reduced the overall variation in hospital readmission rates by about 10 percent and would cut penalties by roughly 40 to 50 percent for the hospitals most affected.²⁷ Safety-net hospitals that serve the sickest, poorest populations tend to absorb the heaviest penalties, not necessarily because their care is worse but because their patient mix is harder. Risk scores used for penalty calculations and risk scores used at the bedside face different challenges, but both are limited by the same underlying truth: the data available at the time of prediction does not fully capture the reasons patients come back.

Methodological Pitfalls That Affect Every Score

Two technical problems lurk beneath every readmission score and rarely get discussed outside the research community. The first is the challenge of distinguishing planned from unplanned readmissions. If a patient comes back for a scheduled chemotherapy session, that should not count against a hospital. CMS uses an algorithm to flag planned readmissions and exclude them, but validation has shown poor accuracy for some of the most common procedures. Diagnostic cardiac catheterizations were correctly identified as planned only about 25 percent of the time, and cardiac device procedures only about 33 percent.²⁸ When planned readmissions are miscounted as unplanned, scores and penalty rates both get noisier.

The second problem is competing risks, particularly death. A patient who dies shortly after discharge cannot be readmitted, and standard statistical models do not handle that well. Research using competing risk methods has shown that hospitals with higher post-discharge mortality can appear to have lower readmission rates under standard models, a distortion that reverses the intended signal.²⁹ A hospital could theoretically look better on readmission metrics by having sicker patients who die rather than return, which is obviously not a sign of quality.

Why Implementation Remains Difficult

Even when a model performs well in validation studies, getting it into clinical workflows is another matter. A qualitative study at one health system documented the barriers firsthand: some care teams had no formal readmission assessment process at all, others were attached to existing tools like LACE and resisted switching, and some were unaware that a newer model had already been built and was available. Clinicians raised concerns about whether the model’s predictors were relevant to their daily decisions, the effort required to change workflows, and the risk of unintended consequences, like over-referring patients to home health services and stretching already thin community resources.³⁰

This gap between model development and clinical use is the field’s least-discussed but most consequential problem. Hundreds of papers describe new readmission models, but relatively few describe what happens when a hospital actually tries to use one to change care. A score that no one looks at, or that identifies patients too late for the care team to act, has zero clinical value regardless of its c-statistic. The evidence increasingly points toward a system where risk prediction is automated, running continuously in the background of electronic health records, flagging patients at admission and again at discharge, and feeding directly into transition-of-care workflows rather than existing as a number a clinician has to look up and interpret on their own.