Understanding NNT and NNH

Essential Tools for Patient-Centered Healthcare and AI Decision Support

A deeper exploration of how these metrics can transform shared decision-making in the age of artificial intelligence

Throughout my journey in chronic care advocacy and AI in healthcare, one concept has consistently emerged as both a clinical metric and a transformative mindset: the Number Needed to Treat (NNT) and its companion, the Number Needed to Harm (NNH). As someone navigating both personal health challenges and the evolving landscape of healthcare AI, I've come to understand these are not just statistical tools. They are bridges between complex medical evidence and meaningful human conversations about care.

The growing integration of AI in healthcare makes understanding these metrics even more critical. As we develop AI systems that can process vast amounts of medical literature and provide decision support, we need frameworks that preserve the human element in medical decision-making while leveraging computational power to make complex information accessible.

What Are NNT and NNH?

The Number Needed to Treat (NNT) tells us how many patients need to receive a treatment for one person to benefit compared to not receiving treatment. An NNT of 25 means that 25 people need to take a medication for one person to avoid a bad outcome, such as a heart attack, that would have occurred without treatment. Lower NNT values indicate more effective treatments.

The Number Needed to Harm (NNH) shows how many patients would need to receive a treatment for one person to experience a harmful side effect. An NNH of 100 means that roughly 1 in 100 patients might experience a serious adverse reaction. Higher NNH values are preferable, indicating that harm is less likely.

The Critical Role of Confidence Intervals

What many discussions of NNT/NNH miss is the importance of confidence intervals (CIs), which reveals the precision and reliability of these estimates. A confidence interval provides a range within which the true value is likely to exist, offering crucial context for interpretation.

For example, if a statin has an NNT of 25 with a 95% confidence interval of 15 to 40, we can be reasonably confident that the true NNT falls somewhere in that range. When treatment effects are not statistically significant, confidence intervals become particularly complex, sometimes encompassing seemingly contradictory regions.

The British Medical Journal now requires that NNTs be reported with 95% confidence intervals whenever possible in randomized controlled trials, recognizing that a wide confidence interval indicates an imprecise outcome that warrants caution in interpretation, regardless of statistical significance.

Understanding the importance of confidence intervals: Think of Confidence Intervals as Your "Margin of Error"

When we say a medication has an NNT (Number Needed to Treat) of 25, we're essentially saying "we need to treat 25 patients to help one person." However, here's the critical clinical insight: that number isn't carved in stone. The confidence interval tells us how reliable that estimate really is.

A Case Study of NNT

Using a statin example: NNT of 25 with a confidence interval of 15-40 means:

• Best case scenario: You might only need to treat 15 patients to help one (much more cost-effective)
• Worst case scenario: You might need to treat 40 patients to help one (significantly more expensive)
• The reality: It's probably somewhere in between, but we can't pinpoint exactly where

Why This Matters for AI Healthcare Development

This is where your future in AI healthcare gets exciting, Dan! AI systems need to handle this uncertainty intelligently. When confidence intervals are wide (indicating uncertainty), AI decision-support tools should flag this for clinicians rather than presenting a false sense of precision.

The BMJ's requirement for reporting these intervals reflects a growing recognition that precision matters as much as the result itself. In AI healthcare, this translates to building systems that are transparent about uncertainty - a critical feature for gaining physician trust and ensuring patient safety.However, here's the critical business insight:

The Science of Risk Communication: Why Format Matters

The power of NNT and NNH extends beyond their numerical values to how they transform our ability to communicate medical information effectively. This connects directly to groundbreaking research in cognitive psychology and medical decision-making.

Dr. Gerd Gigerenzer's research at the Max Planck Institute has demonstrated that people (including physicians) make dramatically better medical decisions when risk information is presented as "natural frequencies" rather than probabilities or percentages. Studies show that natural frequencies help even elderly patients and those with low numeracy skills better understand medical screening test results.

The implications for AI-powered decision support systems are profound. Major evidence-based medical societies, including the Cochrane Collaboration, the International Patient Decision Aid Standards Collaboration, and regulatory agencies, now recommend using natural frequencies because they align with how our minds naturally process risk information.

Evidence from Medical Decision-Making Research

Research published in Psychological Science in the Public Interest demonstrates that when medical information is presented using natural frequencies, physicians' diagnostic accuracy improves significantly, with most physicians accurately estimating positive predictive values. This has direct applications for how AI systems should present treatment recommendations.

Studies in Medical Decision Making and Patient Education and Counseling consistently show that patients make more informed choices when risk is presented in formats that reduce cognitive load and align with natural reasoning processes. This research foundation supports the integration of NNT/NNH into both clinical practice and AI-powered decision support tools.

Real-World Clinical Example: Statins Revisited

Consider statins for cardiovascular prevention, a decision millions face annually. Recent meta-analyses suggest:

• Primary prevention NNT: Approximately 100-300 over 5 years (95% CI: 75-400)
• Secondary prevention NNT: Approximately 25-50 over 5 years (95% CI: 20-75)
• NNH for muscle-related side effects: Approximately 100-250 (95% CI: 80-350)
• NNH for new-onset diabetes: Approximately 200-500 (95% CI: 150-750)

These confidence intervals reveal important nuances. The wide ranges suggest that we need further research, and the overlapping intervals for different populations highlight the importance of personalized medicine. An ideal intervention would have a single-digit NNT for efficacy and a double-digit or higher NNH for adverse outcomes.

Shared Decision-Making: Where Science Meets Humanity

The real power of NNT and NNH emerges in shared decision-making (SDM), a collaborative process in which healthcare choices are made by patients and healthcare professionals working together, with information shared and patients supported in expressing their preferences.

The concept of shared decision-making was first introduced by the 1982 President's Commission for the Study of Ethical Problems in Medicine, which recognized patients as key partners in the decision-making process. However, despite decades of research showing its benefits, SDM has not yet been widely adopted in practice.

Patient decision aids, which often incorporate NNT and NNH data, are essential tools that "provide information on options and help people think about, clarify and communicate the value of each option to them personally". These tools do not recommend specific choices but rather facilitate informed decision-making.

AI and the Future of Evidence-Based Care

As we develop AI systems for healthcare, the principles underlying NNT and NNH become even more crucial. AI excels at processing vast amounts of clinical trial data, but the challenge lies in presenting this information in ways that support rather than replace human judgment.

Recent research shows that despite clear guidelines, only 5 of 88 articles in leading medical journals mentioned NNT or NNH values, even though they could have been calculated for most primary outcomes. AI systems could help bridge this gap by automatically calculating and presenting these metrics with appropriate confidence intervals.

The COVID-19 pandemic demonstrated both the potential and challenges of rapid decision aid development. Researchers successfully created decision aids using digital platforms and asynchronous communication, suggesting new models for AI-assisted shared decision-making.

Limitations and Considerations

NNT and NNH are not without limitations. They can vary substantially over time and may convey different information depending on when they are calculated. They also assume baseline risks and time horizons that may not apply to individual patients.

The calculation and interpretation of NNT depend heavily on study characteristics, including design and outcome variables, and baseline risks, clearly defined outcomes, time horizons, and confidence intervals should always be provided.

Practical Applications for Healthcare AI

For those of us working at the intersection of AI and healthcare, several principles emerge:

1. Transparency: AI systems should present both the numbers and their limitations
2. Context: Baseline risks and time frames must be clearly communicated
3. Personalization: Tools should help adapt population-level data to individual circumstances
4. Cognitive accessibility: Information should be presented in formats that align with natural human reasoning

Looking Forward: A Personal and Professional Perspective

Living with a chronic condition while working in healthcare AI has taught me that the most sophisticated algorithms mean nothing if they cannot help real people make better decisions about their care. NNT and NNH offer a framework for preserving human agency while leveraging computational power.

As Gigerenzer and colleagues observe, "Statistical literacy is a necessary precondition for an educated citizenship in a technological democracy". In an era where AI systems increasingly influence medical decisions, our responsibility is to ensure these tools enhance rather than diminish shared decision-making.

The future of healthcare AI is not about replacing physician judgment or patient preferences. It is about creating systems that make complex evidence more accessible, understandable, and actionable for everyone involved in healthcare decisions.

Sources and Further Reading

Core Methodological References

• Altman, D.G. (1998). Confidence intervals for the number needed to treat. BMJ, 317(7168), 1309-1312.
• Cook, R.J., & Sackett, D.L. (1995). The number needed to treat: a clinically useful measure of treatment effect. BMJ, 310(6977), 452-454.

Risk Communication and Cognitive Psychology

• Gigerenzer, G., Gaissmaier, W., Kurz-Milcke, E., Schwartz, L.M., & Woloshin, S. (2008). Helping doctors and patients make sense of health statistics. Psychological Science in the Public Interest, 8(2), 53-96.
• Hoffrage, U., & Gigerenzer, G. (1998). Using natural frequencies to improve diagnostic inferences. Academic Medicine, 73(5), 538-540.
• Galesic, M., Gigerenzer, G., & Straubinger, N. (2009). Natural frequencies help older adults and people with low numeracy to evaluate medical screening tests. Medical Decision Making, 29(3), 368-371.

Shared Decision-Making Research

• Stacey, D., et al. (2017). Decision aids for people facing health treatment or screening decisions. Cochrane Database of Systematic Reviews, 4, CD001431.
• Elwyn, G., et al. (2012). Shared decision making: a model for clinical practice. Journal of General Internal Medicine, 27(10), 1361-1367.

Contemporary Applications

• Newell, C., et al. (2020). Availability and use of number needed to treat (NNT) based decision aids for pharmaceutical interventions. Patient Education and Counseling, 103(4), 834-841.
• Silva, V., et al. (2017). Number needed to treat (NNT) in clinical literature: an appraisal. BMC Medicine, 15, 112.

Digital Resources

• TheNNT.com - Evidence-based summaries with NNT/NNH data
• Ottawa Hospital Research Institute - Patient Decision Aid Library
• International Patient Decision Aid Standards (IPDAS) Collaboration resources

Executive Abstract: How NNT and NNH Metrics Can Humanize AI in Clinical Practice

In the evolving intersection of AI and chronic care, metrics like Number Needed to Treat (NNT) and Number Needed to Harm (NNH) serve as more than clinical statistics — they are tools for shared decision-making and ethical communication.🔍 

Understanding NNT and NNH at 30K Feet

• NNT indicates how many patients must be treated for one to benefit (lower is better).
• NNH shows how many must be treated before one experiences harm (higher is safer).
• Confidence intervals (CIs) provide the critical context — revealing the range of possible outcomes and uncertainty in data, which AI systems must surface, not obscure.

Why These Metrics Matter in AI Development

• AI can rapidly synthesize data, but must communicate uncertainty clearly.
• BMJ now requires NNTs to include 95% CIs in trials — a signal that precision and transparency are essential.
• AI tools should flag imprecision, not mask it with false certainty.

The Psychology of Risk Communication

• Research by Dr. Gerd Gigerenzer shows that people (including doctors) better understand medical risks when data is shared as natural frequencies, not percentages.
• Presenting NNT/NNH this way improves patient understanding, physician accuracy, and treatment choices.
• This has major implications for how AI systems should format and deliver medical evidence.

Shared Decision-Making (SDM): Science Meets Humanity

• Despite strong evidence, SDM remains underused.
• Patient decision aids that include NNT/NNH help patients weigh options based on values and preferences, not just clinical stats.
• AI can power these tools by making data transparent, contextual, and personally relevant.

Design Principles for AI in Healthcare

1. Transparency — present both results and uncertainty
2. Context — clarify time horizons and baseline risks
3. Personalization — adapt population data to individuals
4. Cognitive accessibility — align output with natural reasoning