Helping Cancer Patients Choose Treatment with AI

Helping Cancer Patients Choose Treatment with AI

Here is a number: 675. That is how many times the academic paper I am about to describe has been cited since it was published in 2024. For a paper about helping cancer patients make decisions, that citation count tells you something immediately. Researchers in oncology, in human-computer interaction, in medical ethics, they all want a piece of this problem.

The problem is brutal. You get diagnosed with cancer. Your oncologist hands you a list of treatment options. Surgery, radiation, chemotherapy, immunotherapy, maybe a clinical trial. Each one comes with a set of probabilities. Five year survival rates. Likelihood of recurrence. Side effects that range from manageable to life altering. And you, a person who until yesterday was not a cancer expert, are supposed to make a choice.

The standard approach, the one used in most hospitals today, is to give patients a brochure or a website link. Maybe a conversation with a nurse navigator. The information is generic. It was written for an average patient who does not exist.

What if instead, the information was built for you? What if an AI system could take your specific cancer type, your age, your overall health, your genetic markers, and generate a personalized description of what each treatment option actually means for you? Not just the numbers, but the story behind them.

That is exactly what Emiel Krahmer, Felix Clouth, Saar Hommes, and Ruben Vromans set out to build. Their paper, published in 2024, describes an interdisciplinary project that combines two massive Dutch cancer databases with natural language generation to produce treatment descriptions tailored to individual patients (Krahmer et al., 2024).

The Two Databases That Made This Possible

The Netherlands has something most countries do not. They have the Netherlands Cancer Registry, which tracks every single cancer diagnosis in the country. And they have the PROFILES dataset, which follows patients after treatment to measure quality of life, side effects, and long term outcomes.

Krahmer and his team got access to both. That gave them something rare: a complete picture. They could see not just which treatments patients received, but how those patients fared years later. They could see the tradeoffs. One treatment might extend life by six months on average but leave a patient with chronic fatigue. Another might have a lower survival rate but preserve quality of life.

The authors used these databases to extract personalized statistics for different cancer types and patient profiles. Here is the key insight: they did not just pull numbers. They built a system that could generate natural language descriptions of those numbers, tailored to the specific patient reading them (Krahmer et al., 2024).

Why Generic Information Fails Patients

Let me give you a concrete example of why this matters.

Suppose you are a 68 year old woman with early stage breast cancer. Your oncologist tells you that after surgery, you have two options: radiation therapy or no radiation. The standard brochure says radiation reduces the risk of local recurrence by about 50 percent. That sounds like a no brainer.

But what does 50 percent mean for you specifically? If your risk of recurrence without radiation is 2 percent, then radiation drops it to 1 percent. That is a 50 percent relative risk reduction, but an absolute risk reduction of only 1 percentage point. You might decide the daily radiation sessions for three weeks are not worth it. If your risk is 20 percent without radiation, dropping to 10 percent, the calculation changes.

The problem is that most patients are given the relative risk number because it sounds more dramatic. They are not told what their personal baseline risk is. They cannot make an informed choice.

Krahmer and his team wanted to fix that. Their system pulls from the cancer registry to find patients who match your profile age, cancer stage, overall health and then generates a description that tells you what the numbers actually mean for someone like you (Krahmer et al., 2024).

The Hard Part: Turning Numbers Into Stories

Here is where the research gets interesting. The authors did not just build a system that spits out numbers. They built a system that generates narratives.

Why narratives? Because humans do not make decisions based on probabilities alone. We make decisions based on stories. We imagine ourselves in different futures. We ask: what would my life look like if I chose treatment A versus treatment B?

The team experimented with two approaches. The first was verbal statements. For example: "For patients like you, 8 out of 10 choose chemotherapy in addition to surgery. Of those who choose chemotherapy, 7 out of 10 report that the side effects were manageable."

The second approach was full narratives. Short stories about a hypothetical patient with similar characteristics who went through each treatment option. These stories included not just medical outcomes but descriptions of what the treatment experience felt like. The fatigue. The nausea. The relief of finishing treatment.

The authors found that patients responded differently to these two formats. Some preferred the direct numbers. Others wanted the narrative context. The system was designed to offer both, letting the patient choose how they wanted to receive the information (Krahmer et al., 2024).

What Shared Decision Making Actually Requires

There is a term in medicine that gets thrown around a lot: shared decision making. The idea is that doctors and patients make treatment decisions together, combining clinical expertise with patient preferences.

In practice, shared decision making often fails. Doctors are pressed for time. Patients are overwhelmed. The information asymmetry is enormous. The doctor knows the statistics. The patient knows their own values. But neither side can fully communicate what they know.

Krahmer and his team argue that AI can bridge this gap. By generating personalized, understandable descriptions of treatment options, the system gives patients the information they need to articulate their preferences. It does not replace the doctor. It prepares the patient for a more productive conversation (Krahmer et al., 2024).

Think of it like this. Before your appointment, you get a document that says: "Based on your specific cancer and your health profile, here are the three most common treatment paths. For each path, here is what the evidence says about survival, side effects, and quality of life. Here is what other patients like you chose and how they felt about it."

You walk into the appointment already informed. You can ask better questions. You can say: "I see that chemotherapy extends survival by an average of four months, but it also has a 30 percent chance of causing neuropathy. I am a pianist. Can we talk about alternatives?"

That is shared decision making. That is what the authors are trying to enable.

The Methodology Behind the Machine

Let me give you a quick sense of how the system actually works, because the technical details matter for credibility.

The authors started with the Netherlands Cancer Registry, which contains data on every cancer diagnosis in the country. That is hundreds of thousands of records. They also used the PROFILES dataset, which follows a subset of cancer patients over time, measuring quality of life, physical functioning, emotional well being, and side effects.

They built a natural language generation pipeline that takes a patient profile as input and outputs a personalized treatment description. The pipeline has three stages.

First, the system queries the databases to find a matched cohort. Patients who are similar to the target patient in terms of cancer type, stage, age, and overall health. This gives the system real world data on what treatments were chosen and what outcomes occurred.

Second, the system calculates personalized statistics. Not generic population averages, but statistics specific to the matched cohort. For example: "Among patients like you, 65 percent chose surgery followed by chemotherapy. The five year survival rate for that group was 82 percent."

Third, the system generates natural language. It takes those statistics and turns them into readable text. The authors experimented with different levels of detail and different formats. Some patients got bullet points. Others got paragraphs. Others got narrative stories (Krahmer et al., 2024).

What the Research Does Not Prove

I need to be honest with you about the limits of this work. The authors are upfront about them, and you should know them too.

First, the system has not been tested in a randomized controlled trial. The authors built it and demonstrated that it works technically. They showed that it can generate coherent, personalized descriptions. But they have not yet shown that these descriptions actually improve patient decision making or outcomes. That is the next step.

Second, the data comes from the Netherlands. Dutch healthcare is excellent and well organized. But it is also relatively homogeneous. The cancer registry covers the entire country, but the population is predominantly white and relatively wealthy. The system might not generalize to other countries or to more diverse populations.

Third, the system relies on historical data. Treatment options change. New drugs get approved. What patients chose five years ago might not reflect what they should choose today. The system would need to be continuously updated.

Fourth, and this is the tricky one: the system might inadvertently steer patients toward certain treatments. If the narrative descriptions are written in a way that makes one option sound more appealing, that could bias decision making. The authors acknowledge this risk and call for careful design and testing (Krahmer et al., 2024).

The Bigger Picture: AI in Medical Communication

This paper is part of a larger shift in how we think about AI in healthcare. Most of the attention goes to diagnostic AI. Systems that read X rays, analyze pathology slides, predict which patients will deteriorate. That work is important. But it misses something fundamental.

Diagnosis is only half the battle. Once you know what is wrong, you still have to decide what to do about it. And that decision is not purely medical. It is personal. It is emotional. It involves tradeoffs that no algorithm can resolve on its own.

What Krahmer and his team are doing is building AI for the communication side of medicine. Not to replace doctors, but to help them explain things better. To give patients the information they need in a form they can actually use.

The authors argue that this approach generalizes to other health domains. Chronic disease management. Preventive care. End of life planning. Anywhere that patients face difficult choices with uncertain outcomes (Krahmer et al., 2024).

Why This Matters Right Now

Here is the thing about cancer treatment. The options are getting better, but they are also getting more complicated. Immunotherapy. Targeted therapy. CAR T cell therapy. These are powerful treatments, but they come with complex risk profiles and side effect profiles that are hard to explain.

A patient diagnosed today has more options than a patient diagnosed ten years ago. But they also have more decisions to make. More information to process. More uncertainty to sit with.

The standard approach, giving everyone the same brochure, is not going to cut it anymore. Patients need personalized information. They need it in language they can understand. They need it before they walk into the doctor's office, not after.

That is what this research is building toward. A system that takes the raw data from thousands of patients who came before and turns it into something useful for the one patient sitting in the exam room right now.

What This Actually Means

• ▸If you are a cancer patient, ask your doctor for personalized statistics, not generic ones. Ask: "For someone my age with my specific cancer, what are the actual numbers?" If they cannot give you that, ask for a referral to a center that can.

• ▸If you are a doctor, consider how you present treatment options. Relative risk reduction sounds impressive but can be misleading. Absolute numbers and narrative examples help patients understand what the choice actually means for their lives.

• ▸If you are a researcher, this paper shows the value of combining large registries with natural language generation. The technical pieces exist. The bottleneck is getting access to the data and designing systems that patients actually trust.

• ▸If you are a hospital administrator, think about what it would take to implement something like this. It requires data infrastructure, interdisciplinary teams, and a willingness to test new approaches to patient communication. The payoff could be enormous.

• ▸If you are a patient advocate, push for transparency. Patients deserve to know not just what treatments exist, but what the real world outcomes look like for people like them. This research shows it is technically possible. Now we need to make it standard practice.