How Employers Can Reduce Avoidable High-Cost Claims With Evidence-Linked AI

From “black box” predictions to traceable, clinically grounded action 

Every self-insured employer has a version of the same story: 

• A single NICU stay, cancer episode, MSK surgery cascade, or ICU admission blows through stop-loss. 

• When you dig in, you see years of warning signs—missed follow-ups, uncontrolled diabetes or hypertension, untreated depression, escalating MSK pain. 

• None of it surfaced early enough in a way your team could act on. 

So you’re told: “Use AI to predict high-cost claimants.” 

That sounds good on paper—but a lot of current “AI” is: 

• A black-box score on a dashboard 

• Built mostly on claims, not real clinical evidence 

• Hard to explain to employees, clinicians, or regulators 

The emerging alternative is evidence-linked AI: systems that don’t just flag risk, but show their work—linking each prediction to specific, traceable clinical evidence across claims, labs, and records. 

This article covers: 

1. Why traditional “high-cost prediction” falls short 

2. What evidence-linked AI actually means 

3. The kinds of high-cost claims that are truly avoidable 

4. How to design an evidence-linked AI program that employers can trust 

5. Guardrails to keep AI aligned with employees’ health and privacy

---

1. Why traditional high-cost prediction isn’t enough 

Risk models for “next year’s high-cost claimants” have been around for years. The research is clear on both their promise and limits: 

• Models based purely on claims data can identify a portion of future high-cost patients, but performance is modest and highly dependent on data quality and chosen features. 

• Adding clinical data from EHRs—labs, vitals, diagnoses, medications—improves predictive performance over claims-alone models for hospitalization, cost, and utilization. 

• Even advanced machine learning models (gradient boosting, random forests, deep learning) can be hard to interpret and may embed existing biases if trained on unexamined administrative data. 

In other words: 

A black-box “risk score” based on messy data is not the same as an actionable, trustworthy early warning system. 

From an employer perspective, three problems show up: 

1. You can’t see why someone is flagged (hard to explain to employees and clinicians). 

2. You can’t tell which flags are truly actionable (e.g., rising-risk MSK vs late-stage metastatic cancer). 

3. You can’t audit or defend the model if someone challenges fairness, privacy, or accuracy. 

That’s where evidence-linked AI changes the game. 

2. What “evidence-linked AI” really means

 

Evidence-linked AI is built on a simple expectation: 

Every prediction or flag must be supported by concrete, inspectable evidence from the member’s longitudinal data. 

That means: 

• Combining multiple data sources—claims, pharmacy, labs, problem lists, sometimes unstructured notes—into a longitudinal picture (what some FDA and ISPOR guidance call “fit-for-purpose real-world data”). 

• Using AI/ML to identify patterns associated with future high-cost events (e.g., rapidly worsening A1c, repeated ED visits for chest pain, escalating opioid fills + MSK imaging). 

• For each flagged member, generating a short, structured “evidence card”, such as: 

• “3 ED visits in 6 months for uncontrolled asthma” 

• “A1c rose from 7.4 to 9.2 in 12 months; no follow-up in 9 months” 

• “Multiple opioid prescriptions + spine MRI + PT drop-off → high-risk MSK trajectory”

This approach aligns with evolving best practices for real-world data: 

• FDA guidance stresses the need for data provenance, traceability, and transparency when using EHR and claims data for decision-making. 

• JAMA Network Open work on data reliability (accuracy, completeness, traceability) shows that combining multiple sources and surfacing traceable evidence dramatically increases reliability. 
  

For employers, the benefit is straightforward: you’re not just told “this person is high risk,” but why, and what you might do about it. 

3. Which high-cost claims are actually avoidable? 

Not every high-cost claim is preventable. But a meaningful slice is foreseeable and modifiable. 

3.1 Cardiometabolic “slow-motion” disasters 

Evidence from diabetes and cardiovascular research shows: 

• Poorly controlled diabetes substantially increases risk of hospitalization, cardiovascular events, and long-term disability—and drives much higher medical costs and productivity losses. 

• Interventions that improve glycemic and blood pressure control reduce hospitalizations and long-term complications, particularly when applied early. 
  

Evidence-linked AI can detect: 

• Rising A1c without treatment intensification 

• Gaps in medication adherence 

• Repeated urgent visits for chest pain, dyspnea, or hypertensive urgency 

These are classic “rising risk” patterns that often precede high-cost cardiac and stroke events. 

3.2 MSK spirals and avoidable surgeries 

Musculoskeletal (MSK) conditions are top cost drivers for employers, and research highlights: 

• Substantial variation in treatment pathways and surgery rates not explained by underlying need. 

• Early, guideline-concordant conservative care (PT, exercise, behavioral support) can reduce surgery rates and total costs. 

AI can flag: 

• Recurrent back or joint pain visits 

• Escalating imaging and opioid use without adequate PT 

• Sudden shifts toward invasive procedures 

These signals enable early outreach and redirection to conservative, evidence-based pathways before a costly, often low-value surgery is scheduled. 

3.3 Decompensating mental health and SUD 

Workforce mental health has become a dominant concern: 

• Large employer surveys show depression, anxiety, and substance use as major drivers of both cost and productivity loss. 

• Studies link untreated depression and SUD with higher risk of ED visits, medical complications, and disability. 

Early warning patterns include: 

• Increased utilization of urgent care / ED for nonspecific complaints 

• Rising prescriptions for sedatives, anxiolytics, or opioids 

• Drop-off from ongoing therapy or psychiatric follow-up 

Evidence-linked AI can surface these patterns in time to offer confidential, voluntary support—EAP, virtual mental health, or integrated behavioral care—before crises generate both human and financial cost. 

4. How employers can design an evidence-linked AI program (without being data scientists) 

You don’t need to build a research lab. But you do need to set the rules of the game. 

4.1 Anchor on three principles 

When you evaluate or design AI solutions, look for: 

1. Multi-source, clinically grounded data 

2. Uses claims and clinical data (labs, vitals, diagnoses, pharmacy) where possible, consistent with RWD best practices. 

3. Evidence cards, not just scores 

4. For each flag, the system should show plain-language rationale plus specific, traceable data points. 

5. Actionable routing, not just risk ranking 

6. Flags should map to clear next steps: nurse outreach, care management, mental health support, second-opinion pathways, etc. 
   

4.2 Start with one or two high-yield cohorts 

Rather than “AI for everything,” start with: 

• Cardiometabolic risk (diabetes, hypertension, heart failure) 

• MSK and chronic pain 

The literature supports strong links between early intervention and reduced hospitalization, surgery, and cost in these domains. 

Define: 

• The patterns you want AI to watch for (e.g., rising A1c without follow-up, multiple back-pain visits + imaging). 

• The programs you already pay for (disease management, telehealth, PT, digital MSK, mental health) that can receive those referrals. 

4.3 Formalize “evidence-linked” workflows 

Design processes so that: 

• Care managers see the evidence alongside the risk flag when they open a case. 

• Your team can audit cases: pull 5–10 flagged members each quarter and verify that the evidence is real and the outreach was appropriate. 

• You can generate simple outcome metrics: 

• Hospitalizations per 1,000 among flagged vs similar unflagged members 

• Surgery rates and costs for MSK 

• Adherence and control metrics in cardiometabolic cohorts 

This echoes how real-world evidence studies structure sensitivity analyses and outcome evaluation using linked claims and EHR data. 

5. Governance & guardrails: keeping AI aligned with employees 

Every employer should put guardrails around how AI touches member data. 

Evidence and guidance from FDA, ISPOR, and health policy literature stress: 

• The need for transparency, data provenance, and fairness checks in AI systems built on real-world data. 

• The risk that algorithms can encode existing inequities in access and treatment if not properly evaluated. 

For employers, practical steps include: 

• Clear use boundaries 

• AI-generated risk flags are used to offer support and care, not to deny coverage, penalize individuals, or make employment decisions. 

• De-identification and “need-to-know” access 

• Limit which internal stakeholders can see identifiable health data; most HR and leadership should only see aggregated trends, not individual cases. 

• Independent review of models 

• Ask vendors to provide documentation of training data sources, performance, and bias checks (e.g., performance across age, gender, race where data supports it). 

• Employee communication 

• Explain, in plain language, that: 

• You use data and AI to spot patterns and offer help earlier, 

• Participation in specific programs is voluntary, 

• Data is protected under HIPAA/ERISA and not used for disciplinary actions. 

This aligns with the direction of both regulatory thinking and public expectations around AI in healthcare. 

6. What success looks like for an employer 

A well-run, evidence-linked AI program doesn’t eliminate high-cost claims—but it changes the mix: 

• Fewer catastrophic events that were clearly foreseeable in hindsight. 

• More members getting earlier, guideline-consistent care for chronic conditions and MSK. 

• Better mental health and SUD support before crises. 

• Clearer stories and metrics you can bring to finance and leadership: 

• “Among flagged rising-risk diabetics who engaged with outreach, we saw 25% fewer hospitalizations and $X per member lower spend over 18 months.” 

• “In MSK, we reduced surgical rates by Y% among members who received early PT + second opinions.” 
  

And crucially: 

• You can show exactly which evidence triggered each intervention. 

• You have governance documents and audits that demonstrate your AI is being used to improve care, not to quietly ration it. 
  

In a world where healthcare costs keep climbing and AI hype is everywhere, that combination—early warning, real evidence, and human-centered guardrails—is what separates employers who genuinely bend their trend from those who just change vendors every renewal. 

References  

1. Kansagara D, et al. Risk Prediction Models for Hospital Readmission: A Systematic Review. JAMA. 2011. 

2. Bertsimas D, et al. Predicting Future High-Cost Patients: A Machine Learning Approach. Health Care Manag Sci. 2020. 

3. Obermeyer Z, Emanuel EJ. Predicting the Future — Big Data, Machine Learning, and Clinical Medicine. N Engl J Med. 2016. 

4. Ross MK, et al. Adding Clinical Data to Claims-Based Risk Models Improves Prediction of Costs and Utilization. Med Care. 2019. 

5. Zeltzer D, et al. Prediction Accuracy With Electronic Medical Records Versus Administrative Claims. Med Care. 2019. 

6. Shadmi E, et al. Predicting 30-Day Readmissions With Claims and Clinical Data. Med Care. 2015. 

7. Riskin DJ, et al. Implementing Accuracy, Completeness, and Traceability for Data Reliability in Real-World Data. JAMA Netw Open. 2025. 

8. FDA. Real-World Data: Assessing Electronic Health Records and Medical Claims Data to Support Regulatory Decision-Making for Drug and Biological Products. Guidance for Industry. 2023–2024. 

9. Franklin JM, et al. Real-World Evidence for Medical Product Evaluation: Challenges and Opportunities. Clin Pharmacol Ther. 2019. 

10. Park J, et al. Productivity Loss and Medical Costs Associated With Type 2 Diabetes Among Employees Aged 18–64 Years. Diabetes Care. 2011. 

11. Stratton IM, et al. Association of Glycaemia With Macrovascular and Microvascular Complications of Type 2 Diabetes (UKPDS 35). BMJ. 2000. 

12. Busse JW, et al. Appropriate Use of Interventions for Low Back Pain. Lancet. 2018. 

13. Deyo RA, et al. Overtreating Chronic Back Pain: Time to Back Off? J Am Board Fam Med. 2009. 

14. Wang PS, et al. Effects of Major Depression on Moment-In-Time Work Performance. Am J Psychiatry. 2004. 

15. Vigo D, Thornicroft G, Atun R. Estimating the True Global Burden of Mental Illness. Lancet Psychiatry. 2016. 

16. Hilty DM, et al. Telepsychiatry and Integrated Care: Evidence-Based Practices and Opportunities for Employers. Psychiatr Clin North Am. 2022. 

17. Kahn JM, et al. Design and Implementation of AI in Healthcare: Ensuring Safety, Efficacy, and Equity. JAMA. 2023. 

18. Jain SH. The Overlooked Frontier Of Value-Based Care: Managing The Rarest, Sickest And Most Expensive Patients. Forbes. 2025.