Can Large Language Models Predict Your Next Diagnosis? A Reformation in Clinical AI

Can AI become a reliable partner in real-time clinical decision-making? A recent study explores whether large language models can predict clinical diagnoses and hospital readmissions based on patients’ medical histories.

Introduction

Today, healthcare systems across the globe are navigating the pressure of increasing complexity: aging populations, staff shortages, rising costs, and a wave of chronic disease. At the same time, more and more patient information is being recorded in electronic health records (EHRs) – not only structured data such as diagnoses, lab values, and medications, but also unstructured notes like physician letters. Artificial intelligence, in the form of large language models (LLMs), can use this wealth of data to calculate probabilities of future clinical events. Trained on millions of anonymized patient histories, LLMs learn typical patterns of disease progression and risk factors. When applied in real-world care, such a model can then analyze an individual patient’s record and estimate, for example, whether hospital readmission in the next two weeks is likely or whether a new diagnosis is emerging.

In this way, AI does more than rely on a doctor’s notes or experience. It processes the entire medical history in relation to countless comparable cases and turns abundant but often overwhelming data into actionable foresight [1].

What are the core challenges? Traditional statistical models, like logistic regression, are often unable to interpret the complex, unstructured text records and time-based clinical data [2]
. Additionally, EHR data is noisy, fragmented, inconsistent, and not originally designed for prediction[3–5]. This has led researchers to ask: can LLMs be fine-tuned to provide clinically meaningful predictions based on EHRs – accurate, interpretable, and practically useful?

Clinical Prediction with Large Language Model (CPLLM)

In their 2024 study, Shoham and Rappoport introduced Clinical Prediction with Large Language Model (CPLLM). This method fine-tunes general-purpose LLMs, such as Llama2 and BioMedLM, on real patient histories, with the goal of predicting future medical outcomes [2]. The researchers used large datasets from intensive care units, converted structured data into textual form and trained their models to answer key questions: Will this patient be diagnosed with a particular disease at their next hospital visit? Will they be readmitted within 15 days?

Figure: Illustration of the fine-tuning process for prediction

Unlike older models, CPLLM improves predictive analytics by using LLMs to process structured EHRs. It does not only consider single data points but also the sequence of events in a patient’s history. This ability leads to higher accuracy in forecasting diseases and assessing readmission risks. In evaluations, CPLLM outperformed traditional approaches and can also be adapted to a range of clinical tasks, such as diagnosing chronic conditions or predicting treatment responses, without requiring extensive domain-specific pre-training [2].

Great Performance – Implementation Remains Challenging

Although CPLLM shows strong performance, several challenges remain before it can be used in routine care.

1. Data privacy: Patient data must be handled under strict regulations such as HIPAA in the US and GDPR in Europe. Compliance is essential to protect sensitive information [2, 6, 7].
2. Interpretability: The model often works as a “black box.” If predictions are not transparent, clinicians may hesitate to trust them in critical situations [7, 8].
3. Data quality: Electronic health records are often incomplete, inconsistent, or fragmented. Without standardization, predictions may be biased or unreliable [9].

These challenges show that technical progress alone is not enough. Trust, transparency, and clear frameworks are equally important for safe use in healthcare.

Conclusion and Implications

This article has shown that CPLLM, a large language model fine-tuned on patient histories, can forecast future diagnoses and hospital readmissions more accurately than traditional approaches. What this means in practice is not a replacement of medical judgment, but the addition of a complementary tool that supports decision-making by providing early warnings or second opinions in complex cases. For digital health startups, CPLLM signals a moment of opportunity. The technology provides a ready-to-use engine for predictive features—whether in triage tools, chronic disease apps, or integrated EHR solutions. With open-source code and compatibility with common platforms like Hugging Face, implementation is realistic even for smaller teams. The key is to combine this technology with strong user experience, clinical oversight, and compliance with regulations.

For healthcare professionals, it’s time to become not just users but evaluators of AI. Understanding how a model like CPLLM works—what it can and cannot do—will be critical in building meaningful, human-centered clinical AI workflows. Trust will not come from accuracy alone. It will come from transparency, explainability, and ongoing collaboration between humans and machines.

And for the general public, this research is a glimpse into a future of healthcare that is more predictive, more personal, and more responsive. But it also comes with responsibilities. As AI becomes more integrated into care, questions about data privacy, informed consent, and algorithmic fairness will only grow. Patients must be not just protected, but empowered to understand how AI helps shape their care.

In conclusion, CPLLM offers more than an incremental advance in clinical AI. It presents a vision of healthcare where machines are trained to understand patient histories, and it calls us to rethink how care is planned, delivered, and experienced.

Literature

1. Kim MK, Rouphael C, McMichael J, Welch N, Dasarathy S. Challenges in and Opportunities for Electronic Health Record-Based Data Analysis and Interpretation. Gut Liver. 2024;18:201–8. doi:10.5009/gnl230272.
2. Ben Shoham O, Rappoport N. CPLLM: Clinical prediction with large language models. PLOS Digit Health. 2024;3:e0000680. doi:10.1371/journal.pdig.0000680.
3. Li Y, Rao S, Solares JRA, Hassaine A, Ramakrishnan R, Canoy D, et al. BEHRT: Transformer for Electronic Health Records. Sci Rep. 2020;10:7155. doi:10.1038/s41598-020-62922-y.
4. Li Y, Mamouei M, Salimi-Khorshidi G, Rao S, Hassaine A, Canoy D, et al. Hi-BEHRT: Hierarchical Transformer-Based Model for Accurate Prediction of Clinical Events Using Multimodal Longitudinal Electronic Health Records. IEEE J Biomed Health Inform. 2023;27:1106–17. doi:10.1109/JBHI.2022.3224727.
5. Rasmy L, Xiang Y, Xie Z, Tao C, Zhi D. Med-BERT: pretrained contextualized embeddings on large-scale structured electronic health records for disease prediction. NPJ Digit Med. 2021;4:86. doi:10.1038/s41746-021-00455-y.
6. Tripathi S, Mongeau K, Alkhulaifat D, Elahi A, Cook TS. Large Language Models in Health Systems: Governance, Challenges, and Solutions. Acad Radiol 2024. doi:10.1016/j.acra.2024.09.047.
7. Ullah E, Parwani A, Baig MM, Singh R. Challenges and barriers of using large language models (LLM) such as ChatGPT for diagnostic medicine with a focus on digital pathology – a recent scoping review. Diagn Pathol. 2024;19:43. doi:10.1186/s13000-024-01464-7.
8. Nasarian E, Alizadehsani R, Acharya U, Tsui K-L. Designing interpretable ML system to enhance trust in healthcare: A systematic review to proposed responsible clinician-AI-collaboration framework. Information Fusion. 2024;108:102412. doi:10.1016/j.inffus.2024.102412.
9. Jetley G, Zhang H. Electronic health records in IS research: Quality issues, essential thresholds and remedial actions. Decision Support Systems. 2019;126:113137. doi:10.1016/j.dss.2019.113137.