Predictive Analytics in Healthcare: What It Can Predict, and How Good Models Still Fail

What predictive analytics is

Descriptive analytics tells you what happened: last month’s readmission rate, yesterday’s emergency department wait. Predictive analytics goes one step further and gives a probability for something that has not happened yet, such as “this patient has a 35 percent chance of missing tomorrow’s appointment.”

The technique ranges from a simple scoring formula to machine learning on thousands of variables. The label matters less than the question: what decision will this prediction change? Our guide to business intelligence in healthcare covers the descriptive side that predictive work builds on.

Where it is used

The table lists common prediction targets and what a team might do with the answer. It is our summary of typical uses, not a ranking, and results vary a great deal between hospitals.

Notice that every row ends with a human action. A prediction that nobody acts on has no value. Staffing questions also connect to the broader nursing workforce picture.

How a model gets built

1. Define the decision. Who will act on the prediction, and what will they do differently?
2. Choose the outcome. “Readmitted within 30 days” is clear. “Sick” is not.
3. Collect and clean the data. Records, labs, past visits, and sometimes social factors. Check for gaps and errors.
4. Split the data. Train on one part and test on a part the model has not seen.
5. Evaluate on measures that match the decision. The next section shows why.
6. Test it locally and prospectively. Try it on recent data from your own setting, ideally running quietly alongside normal work before anyone relies on it.
7. Build it into the workflow. A score buried in a separate screen will be ignored.
8. Monitor and update. Patients, coding and practice change over time, and model accuracy can drift.

The accuracy trap: a worked example

Four measures describe how a yes-or-no prediction performs. The NIH’s StatPearls article on diagnostic testing accuracy defines them this way:

• Sensitivity: the proportion of true positives out of all patients with the condition. How many of the real cases does it catch?
• Specificity: the proportion of true negatives out of everyone without the condition. How well does it clear the people who are fine?
• Positive predictive value (PPV): out of all positive findings, how many are true positives. When it flags someone, how often is it right?
• Negative predictive value (NPV): out of all negative findings, how many are true negatives.

The same source notes that disease prevalence in a population affects PPV and NPV. Here is what that means in practice. This is a made-up clinic example.

A clinic books 1,000 appointments. Its no-show model has sensitivity of 80 percent and specificity of 90 percent, which sounds good. Staff will phone every patient the model flags.

What this shows. In Clinic A, the model is 89 percent accurate overall, yet more than half of the people staff phone (90 of 170) would have turned up anyway. The model is the same in both clinics. Only the base rate changed, and PPV moved from 47 to 77 percent. Overall accuracy told you almost nothing about whether the phone calls would be well spent.

The lesson: pick measures that fit the decision. If a missed case is costly, look at sensitivity. If wasted effort is costly, look at PPV and the number of alerts staff must handle. Ask for these numbers for your own patient mix, not the vendor’s demo data.

When models go wrong

Bias from a proxy

A widely cited study by Obermeyer, Powers, Vogeli and Mullainathan, published in Science in 2019 and summarized in the Chicago Booth Review, examined an algorithm used to decide who is offered extra help in high-risk care management programs. The algorithm used future health care costs as a stand-in for health needs. According to the summary, at the same risk score Black patients were sicker than white patients: at the 97th percentile they had 26 percent more chronic illnesses on average. Black patients made up 17.7 percent of those automatically identified for the program, and the researchers calculated the figure would have been 46.5 percent without the bias. Replacing cost with a measure that combined cost predictions with health predictions reduced the bias by 86 percent by one measure.

The takeaway is not that algorithms are always biased. It is that the outcome you choose to predict shapes who benefits, and cost is not the same as need. Check results by group, such as age, sex, race, language and insurance type.

Other common failures

• Drift. A model built before a change in coding, treatment or patient mix may lose accuracy.
• Alert fatigue. Too many low-value alerts teach staff to ignore all of them. This is the PPV problem again.
• Data leakage. A model quietly uses information that would not be available at the moment of prediction, such as a later test result, and looks better in testing than in use.
• Poor transfer. A model that works at one hospital may not work at another. Test locally before relying on a vendor’s published results.
• No owner. If nobody is responsible for maintenance, performance fades unnoticed.

Privacy and rules

Health data is sensitive. In the United States, the HIPAA Privacy Rule governs identifiable health information, and our business intelligence in healthcare guide summarizes the de-identification methods. Other countries and states have their own rules.

Some software that supports clinical decisions may also be regulated as a medical device. The US Food and Drug Administration (FDA) publishes guidance on which clinical decision support software counts as a device, and that guidance has been revised over time, so check the current version and ask your compliance or legal team before deploying a tool that influences care. This guide does not give legal advice.

A rollout checklist

Start with a pilot in one unit, compare outcomes with a control group or earlier period, and expand only if the results hold.

Skills and careers

Working in this area usually combines three skill sets: data skills (SQL, statistics, a language such as Python or R), healthcare knowledge (how care and coding work), and communication (explaining a model to people who will act on it). Roles go by many titles, including healthcare data analyst, data scientist and clinical informatics specialist. See the careers section of our business intelligence in healthcare guide and the business intelligence certificate guide. For another healthcare quality method, see Six Sigma in healthcare.

Common mistakes

• Judging a model by overall accuracy. Look at sensitivity, PPV and alert volume.
• Predicting what is easy to measure, not what matters. Cost is a classic example.
• Skipping the workflow. A score nobody sees changes nothing.
• Trusting vendor results blindly. Test on your own patients.
• Forgetting maintenance. Plan for monitoring from day one.
• Treating a probability as a verdict. A 35 percent risk means 65 times in 100 it will not happen.

Learn more

If you want to build the underlying skills, online providers such as Coursera and Udemy list data analytics and big data courses, some with healthcare examples. We may earn a commission if you sign up, at no cost to you. We are not affiliated with any university. Check the syllabus before you pay, and see all business intelligence guides for related reading.

Frequently asked questions

What is predictive analytics in healthcare?

The use of past data and statistical or machine learning models to estimate the likelihood of future events, such as missed appointments, readmissions or deterioration, so staff can act earlier.

What are examples of predictive analytics in healthcare?

No-show prediction, readmission risk, early warning for deterioration or sepsis, bed demand forecasting, and population risk scoring for care management.

How accurate are healthcare prediction models?

It varies widely by model, setting and measure. Overall accuracy can mislead, so ask for sensitivity, positive predictive value and alert volume tested on your own data.

What is the difference between sensitivity and PPV?

Sensitivity is the share of real cases a model catches. PPV is the share of flagged cases that are real. A model can have high sensitivity and low PPV, which means many false alarms.

Can predictive models be biased?

Yes. A 2019 study found an algorithm that used cost as a proxy for need assigned lower risk to Black patients who were as sick as white patients. Check performance by group.

Is predictive analytics regulated?

Data use is covered by privacy laws such as HIPAA in the US, and some decision support software may be regulated as a medical device. Check current rules with compliance or legal advisers.

Do I need to code to work in healthcare analytics?

Many roles use SQL and a language such as Python or R, though some use no-code tools. Knowledge of healthcare data matters as much as the tools.

How do I start a predictive analytics project in a hospital or clinic?

Pick one decision, define the outcome, measure the baseline, test a simple model on your own data, pilot it in one unit, and keep a named owner for monitoring.