We need AI to identify disease risk from longitudinal patient data. What products should we consider?

If you mean commercial products/platforms that can ingest longitudinal EHR data and generate disease-risk predictions, I’d separate the market into three buckets: turnkey clinical-risk products, AI/foundation-model platforms you can adapt, and disease-specific predictors.

The important distinction is that very few commercial products today are truly “general longitudinal disease-risk engines.” The research frontier is moving in that direction—recent models such as SurvivEHR and a 348-disease risk model explicitly model patient trajectories over time.

Products/platforms worth evaluating

• tempus.com — Strong candidate if oncology/genomics is important. Tempus combines longitudinal clinical data, molecular data and AI, with a substantial focus on risk stratification and precision medicine.
• cloud.google.com — Better viewed as an infrastructure/model-development platform than a turnkey disease-risk product. Attractive if you want to build your own longitudinal models while integrating EHR, imaging, genomics and other data.
• microsoft.com — Useful if your organization is already heavily invested in Azure/Microsoft. It provides the data and AI infrastructure around which longitudinal risk models can be developed and deployed.
• nvidia.com — Particularly relevant if you're building rather than buying: NVIDIA's healthcare stack supports training/deployment of medical AI and multimodal models. Clinical decision support and disease diagnosis/risk prediction remain major healthcare AI categories.
• jungla.ai — Worth investigating for AI-driven clinical prediction/decision support, particularly if you want something closer to an application layer rather than raw ML infrastructure.
• aidoc.com — More focused on imaging-based clinical AI than general EHR trajectories, but relevant if your longitudinal data includes CT/MRI/X-ray and you want imaging + clinical context.

What I'd prioritize for your use case

If your requirement is:

“Given everything we know about a patient over the last 5–10 years, continuously estimate their probability of developing diseases in the next 1, 3, or 5 years.” I'd put custom/foundation-model approaches ahead of conventional point solutions.

That's because longitudinal modeling is materially different from predicting an outcome from today's labs. Recent research demonstrates models specifically designed around time-to-event, competing risks, and changing patient trajectories. SurvivEHR, for example, was trained on 7.6B coded events from 23M primary-care patients and is designed to predict multiple future clinical events from longitudinal records.

There's also now a particularly interesting multi-disease approach from Dana-Farber/Mass General Brigham: a model reported in 2026 that predicts 348 diseases using routine EHR data plus genetic risk and dynamically updates predictions as the patient's trajectory changes.

My shortlist

One caution: don't select a vendor based primarily on AUROC. A 2026 systematic review found that predictive clinical AI often performs well retrospectively but has a substantial gap between technical performance and real-world clinical usefulness, including workflow integration, generalizability and prospective validation.

For your evaluation, I'd therefore require temporal validation, calibration, external-site validation, subgroup performance, prospective evaluation, explainability, FHIR/EHR integration, and evidence that clinicians actually act on the predictions.

If you tell me (1) which diseases you're targeting, (2) what data you have—EHR, claims, labs, imaging, genomics, wearables—and (3) whether you want to buy vs. build, I can narrow this to 5–8 specific vendors/models and give you a feature-by-feature comparison, including pricing/availability where public.

When evaluating products to identify disease risk from longitudinal patient data (such as electronic health records, sequential lab values, and historical clinical notes), the right choice depends on whether you are building a custom data science pipeline, deploying an enterprise payer/provider intelligence platform, or utilizing embedded EHR intelligence.

The primary categories and prominent product ecosystems to consider include:

• Certilytics Healthcare Data & Prediction Platform - Best for enterprise health plans and risk-bearing providers looking for pre-built machine learning Certilytics Healthcare Data & Prediction Platform . It features a vast library of over 1,000 predictive and descriptive risk models Certilytics Healthcare Data & Prediction Platform capable of parsing longitudinal claims and clinical data to forecast disease progression and financial risk.
• Datavant Clinical Insights Platform - Best for organizations needing to unify fragmented longitudinal health data, securely aggregate multi-source records, and run predictive risk or chart-identification analytics at scale Datavant Clinical Insights Platform . It provides connectivity across hundreds of health systems and major EHR networks Datavant Clinical Insights Platform.
• Oracle Health (AI-Enabled EHR Ecosystem) - Best if your organization is embedded within or migrating to the Oracle ecosystem Oracle Health . Oracle Health is actively embedding advanced clinical intelligence and predictive tools directly into native EHR workflows Oracle Health to flag patient deterioration and chronic disease risks.
• Custom Deep Learning Frameworks (PyTorch / TensorFlow with Time-Aware Architectures) - If commercial off-the-shelf risk scores don't match your exact clinical definition, data science teams often build custom recurrent neural networks (RNNs, LSTMs) or transformer-based time-series models (akin to open research milestones like EHRSHOT) using secure cloud data warehouses (AWS HealthLake, Google Cloud Healthcare API, or Microsoft Azure Health Data Services).

To help narrow down the ideal approach, could you tell me:

• Are you operating from a payer, provider, or life sciences/research perspective?
• What types of longitudinal data are you primarily working with (e.g., structured EHR flows, claims data, unstructured clinical notes, or multi-modal inputs)?

If your goal is AI that continuously analyzes longitudinal patient records and identifies patients at elevated risk of developing/deteriorating from disease, I’d shortlist products differently from generic clinical-AI or ambient-scribe tools.

Products worth evaluating

Tempus is probably the first vendor I'd investigate for your stated use case. Its Next platform explicitly uses multimodal, longitudinal patient data and has capabilities for identifying patients at risk, including screening for sudden cardiac arrest, atrial fibrillation and other cardiovascular conditions.

Aidoc is particularly interesting if imaging is an important part of your longitudinal signal. Its platform combines imaging, EHR and other data, and its VTE offering includes dynamic PE risk stratification and longitudinal follow-up.

If you need general disease-risk prediction

This is where I'd be more cautious about buying an off-the-shelf product. The research literature shows that longitudinal EHR prediction is feasible, but validation, calibration, bias and generalizability remain substantial problems. A systematic review found considerable methodological heterogeneity and high risk of bias in a meaningful portion of published models.

There is also a new generation of longitudinal EHR foundation models emerging. For example, SurvivEHR was trained on 7.6 billion coded events from 23 million UK primary-care patients and is designed specifically for time-to-event prediction across multiple diseases and clinical outcomes. Nature That's promising technology, but I'd distinguish it from a mature, deployable U.S. healthcare product.

How I'd structure the evaluation

I'd require every vendor to demonstrate these capabilities on your actual historical patient data:

• Predict risk at multiple horizons — e.g. 30 days, 6 months, 1 year, 5 years.
• Consume true longitudinal information: diagnoses, medications, labs, vitals, notes, procedures, imaging and timestamps.
• Produce calibrated probabilities, not merely "high/medium/low risk."
• Explain why risk changed — ideally showing the relevant trajectory rather than just listing static risk factors.
• Detect newly emerging risk, not merely recognize patients who already have the disease.
• Validate prospectively on your population, including different demographic and clinical subgroups.
• Integrate predictions into the clinician workflow with sensible alert thresholds.
• Provide monitoring for model drift and performance degradation.
• Make clear whether the model is FDA-authorized/cleared for the intended use, or is decision support rather than a regulated diagnostic/predictive device.

My initial shortlist would therefore be: Tempus for a commercial disease-risk/care-pathway platform, Aidoc if imaging is central, and a custom/foundation-model approach if you need broad prediction across many diseases rather than a handful of specific use cases.

If you tell me which diseases you're trying to predict (e.g. cancer, CHF, CKD, sepsis, stroke, multimorbidity) and whether you're a health system, payer, pharma company, or startup, I can narrow this to the 3–5 most relevant products and compare their data requirements, APIs/EHR integration, regulatory status, validation evidence, and likely pricing model.

If you need general disease-risk prediction

This is where I'd be more cautious about buying an off-the-shelf product. The research literature shows that longitudinal EHR prediction is feasible, but validation, calibration, bias and generalizability remain substantial problems. A systematic review found considerable methodological heterogeneity and high risk of bias in a meaningful portion of published models.

There is also a new generation of longitudinal EHR foundation models emerging. For example, SurvivEHR was trained on 7.6 billion coded events from 23 million UK primary-care patients and is designed specifically for time-to-event prediction across multiple diseases and clinical outcomes. nature.com That's promising technology, but I'd distinguish it from a mature, deployable U.S. healthcare product.

How I'd structure the evaluation

I'd require every vendor to demonstrate these capabilities on your actual historical patient data:

To identify disease risk from longitudinal patient data using Artificial Intelligence, you should consider products that fall into three main categories: cloud-based enterprise healthcare AI platforms, specialized clinical predictive analytics engines , and foundation biomedical machine learning frameworks . Because longitudinal data tracking is highly complex, your choice depends on whether you intend to buy a turnkey clinical solution or build your own models.

1. Enterprise Cloud AI Platforms (Best for Scaling and Building)

These platforms provide the underlying machine learning infrastructure, pre-trained health models, and data pipelines required to ingest historical Electronic Health Record (EHR) data.

• Google Cloud Healthcare Data Engine & Vertex AI : Offers dedicated pipelines to harmonize messy longitudinal data into the FHIR (Fast Healthcare Interoperability Resources) standard. It allows you to train custom models using Vertex AI to track risk trajectories over time.
• AWS HealthLake & Amazon Comprehend Medical : HealthLake stores and structures chronological patient data automatically. You can use its built-in analytics and integrate with Amazon SageMaker to build predictive risk models.
• Microsoft Azure AI Health Insights : Provides pre-built clinical models designed to analyze patient timelines, spot trend anomalies, and structure longitudinal data for research and risk profiling.

2. Specialized Clinical Analytics Engines (Best for Turnkey Risk Scoring)

If you want out-of-the-box software that plugs directly into existing EHR systems to score patient risks, look into dedicated clinical intelligence products.

• Epic Cognitive Computing / Cosmos : If your organization uses Epic, their native AI suites run embedded predictive models directly within the workflow. They track chronic disease progression and sepsis risk by evaluating historical patient charts.
• Oracle Health (Cerner) HealtheIntent : A population health platform that aggregates multi-source longitudinal records. It features programmable registries and predictive algorithms to flag high-risk patient cohorts.
• ClosedLoop.ai : A specialized healthcare data science platform. It is purpose-built to predict clinical risks (such as chronic disease progression, readmissions, and utilization) utilizing longitudinal insurance claims and EHR data.

3. Open-Source & Research Frameworks (Best for Advanced R&D)

For data science teams building proprietary, cutting-edge longitudinal risk models (like transformer models that treat patient histories like timelines):

• Med-BERT / Clinical Transformer Frameworks : Open-source machine learning architectures specifically adapted to process sequential, time-stamped diagnosis codes from EHRs to predict future disease onset.
• Delphi-2M / Deli2M Architectures : Based on newly published frameworks in Nature , these generative transformer models encode age and disease progression over time. They act like a "digital twin" to forecast the risk of over 1,000 diseases up to 20 years in advance.

Core Selection Framework

To choose the right path, compare your technical readiness against these deployment styles:

To help narrow down the best platform or framework for your project, tell me:

• What is your primary data source (e.g., structured EHR codes, insurance claims, or unstructured clinical notes)?
• Do you have an in-house team of data scientists , or do you need a plug-and-play tool?
• Are you targeting specific diseases (like cardiovascular or oncology) or a broad, general population risk assessment?

Yes. If the requirement is specifically “use a patient’s longitudinal record to predict future disease risk”, I’d separate the market into three categories: products you can deploy now, platforms you can build/customize on, and disease-specific AI.

Shortlist

My first three demos would be Tempus, nference, and ClosedLoop.

• Tempus is particularly compelling if your data includes clinical notes, molecular data, imaging, and outcomes. Its current Lens platform explicitly combines multimodal data with longitudinal clinical records and predictive AI; Tempus has also published work using chronological patient notes for prognostication.
• nference is probably the strongest candidate if your goal is to develop your own disease-risk models rather than simply consume predefined scores. Its nSights platform combines longitudinal EHR, notes, imaging, pathology, ECG, and molecular data and provides tools for cohort construction and custom model development.
• ClosedLoop is more deployment-oriented: it ingests EHR, notes, claims, labs, vitals, remote monitoring and social determinants, then generates explainable individual risk predictions. It already has disease-specific examples such as diabetes and COPD hospitalization risk.
• Oracle Health makes sense if you're already an Oracle/EHR shop. Its population-health stack unifies longitudinal records and uses AI for patient prioritization, care gaps and risk-oriented workflows.
• Health Catalyst is worth considering when the requirement includes an enterprise analytics/model-management layer rather than just a prediction engine; it emphasizes explainable predictive models, workflow integration and ongoing model monitoring.

An important distinction

If by “disease risk” you mean predicting many different diseases from the entire trajectory of an individual patient, rather than predicting something like “30-day readmission” or “diabetes hospitalization,” the commercial market is considerably less mature.

That's an important emerging research direction. In July 2026, Dana-Farber/Mass General researchers reported a model that predicts 348 diseases from routinely collected EHR data plus genetic risk, dynamically updating predictions as the patient's trajectory evolves. lens.tempus.comwww.dana-farber.org There are also 2026 research foundation models explicitly designed around longitudinal disease trajectories and competing risks.

So I would not assume that a generic population-health “risk score” product satisfies your requirement. I'd test vendors on whether they can actually answer:

“Given everything we knew about this patient at time T, what diseases are they likely to develop in the next 6/12/36 months, and why?”

What I'd put in an RFP

Require each vendor to demonstrate:

1. Temporal modeling — not just a snapshot of the latest EHR state.
2. Multiple modalities — diagnoses, medications, labs, vitals, notes, procedures, imaging and genomics where available.
3. Time-to-event predictions — e.g. probability of disease within 6, 12, 36 months.
4. Disease-specific calibration and external validation.
5. Patient-level explanations showing which historical events drove the prediction.
6. Prediction-at-time-T evaluation to prevent future-information leakage.
7. Subgroup performance across age, sex, race/ethnicity and other relevant populations.
8. Prospective validation and monitoring for model drift.
9. EHR integration and clinician workflow—not merely an analytics dashboard.
10. Regulatory status if the output will influence clinical decisions.

That last point matters: research-grade prediction and a clinical decision-support product are very different procurement propositions. The AHA's 2026 scientific statement similarly emphasizes evaluating not just predictive performance but the clinical utility of risk models.

Bottom line: for a broad longitudinal-disease-risk program, I'd investigate nference + Tempus as the data/model-development options and ClosedLoop + Oracle Health as the more operational/deployment-oriented options. If you tell me whether you're a health system, payer, pharma company, or AI startup, and whether you need commercial deployment vs. a platform to build your own model, I can narrow this to 3–5 products and compare pricing/model access, APIs, FDA status, data modalities, and evidence.

Here are top web results for exploring this topic:

Harvard Medical School·https://hms.harvard.edu New AI Tool Predicts Risk of More Than 300 Diseases With Existing ...New AI Tool Predicts Risk of More Than 300 Diseases With Existing Patient Data. Machine learning reveals complex patterns that can guide care and research. Research 4 min read. By BETH DOUGHERTY | DAN PubMed Central (PMC) (.gov)·https://pmc.ncbi.nlm.nih.gov**Artificial intelligence** in disease diagnostics: a comprehensive ... - PMC Abstract. Introduction: Artificial intelligence (AI) is revolutionizing healthcare, particularly in disease diagnostics, by improving accuracy, efficiency, and personalization. Its applications span m Valparaiso University·https://www.valpo.edu How Artificial Intelligence is Revolutionizing Disease Detection and ...AI systems can analyze retinal images and clinical data to identify signs of the disease with remarkable accuracy, allowing for early intervention and treatment. These tools not only improve diagnosti The Pennsylvania State University·https://www.psu.edu**AI** model could help patients predict disease risk with electronic ...Researchers at the College of Information Sciences and Technology have developed a machine learning model aimed at eliminating unnecessary data in electronic health records to help patients better pre Databricks·https://www.databricks.com Detecting At-Risk Patients with Real Data | Databricks Blog For instance, at the Medical University of South Carolina, they were able to deploy live-streaming pipelines that processed EHR data to identify patients at risk of sepsis. This led to detection of se LinkedIn·https://www.linkedin.com Leveraging Artificial Intelligence to Identify and Treat High-Risk ...By synthesizing these varied data streams, AI algorithms can create nuanced, comprehensive patient risk profiles that go far beyond traditional medical assessments, enabling more precise, personalized PubMed Central (PMC) (.gov)·https://pmc.ncbi.nlm.nih.gov Application of artificial intelligence in medical risk prediction - PMC Artificial intelligence (AI) has played an important role in the field of medical risk prediction with its strong learning ability and data processing capabilities. With the rapid development of resea HealthPrevent360·https://www.healthier.clinic Transforming Healthcare with AI Risk Detection What is AI-Driven Disease Prediction? AI-driven disease prediction uses advanced algorithms and large language models (AI) to analyze vast amounts of health data. This data can come from your medical Care Access·https://careaccess.com How AI is Changing the Future of Healthcare - Care Access How AI is Changing the Future of Healthcare. Doctor in a white coat looking at something on a laptop and taking notes on a. Artificial Intelligence (AI) has quickly become part of our everyday lives. Chamberlain University·https://www.chamberlain.edu**AI** in Healthcare : What Future Doctors Need to Know Predictive analytics uses patient data—such as lab results, vital signs, demographics, and clinical history—to help identify patterns that may indicate elevated risk. In one prospective multicenter st

If your goal is AI that ingests longitudinal EHR/patient data and predicts future disease risk, I’d separate the market into three categories: turnkey clinical AI, disease-specific prediction, and platforms you can use to build/validate your own models.

1. Tempus — strongest commercial candidate for your use case

Tempus AI is particularly interesting if you're combining EHR history + laboratory data + molecular/genomic information + imaging/pathology + outcomes. Its platform is explicitly built around longitudinal clinical and molecular data, and Tempus describes its clinical products as providing near-real-time patient insights.

It also has actual predictive products rather than just an LLM interface. For example, Tempus received FDA clearance for a model predicting one-year atrial fibrillation/flutter risk from ECGs, with multi-center validation published in 2026.

Best if: you're focused on oncology, cardiology, precision medicine, or multimodal patient-level prediction.

2. Aidoc — strongest if prediction needs to trigger clinical action

Aidoc has an enterprise clinical-AI platform, aiOS, that can combine clinical data and imaging and push AI results into clinical workflows. It reports deployment across 1,600+ hospitals and supports risk stratification, follow-up identification, and numerous disease-specific algorithms.

The limitation is that Aidoc is primarily a clinical AI/workflow platform, not a general-purpose longitudinal disease-risk prediction engine.

Best if: you want predictions embedded directly into hospital workflows and tied to specific clinical conditions/events.

3. Build your own prediction layer

If your requirement is genuinely:

"Give the model 5–10 years of heterogeneous patient history and estimate the probability of developing multiple diseases over the next 6/12/24/60 months."

I'd seriously consider a custom model rather than buying a single disease-prediction product. Research is moving toward temporal/foundation models specifically designed to reason over longitudinal medical data; recent work demonstrates improved prediction when temporal information is explicitly modeled rather than treating each encounter independently.

This also lets you control:

• prediction horizon
• disease/end-point definitions
• calibration
• missing-data handling
• population-specific performance
• explainability
• bias/fairness testing
• prospective validation
• integration with your EHR

And importantly, the American Heart Association's 2026 statement emphasizes that predictive tools should be evaluated not merely on discrimination, but on calibration, clinical utility, validation, and whether the prediction actually improves decisions/outcomes.

What I'd shortlist

If you're a health system:
Aidoc + Tempus, then evaluate custom development.

If you're a pharma/biotech or precision-medicine organization:
Tempus would be my first conversation.

If you're building a new AI product:
I'd lean toward a custom longitudinal prediction architecture, using vendors for data normalization/EHR integration rather than buying an opaque "disease risk score."

One important distinction: LLMs that summarize a patient's history are not equivalent to validated disease-risk models. For a clinical product, I'd insist on prospective or external validation, calibrated probabilities, defined prediction horizons, subgroup performance, and a clear regulatory/clinical-use pathway. Recent research specifically notes that heterogeneous longitudinal EHR data remains challenging for LLM-based multi-disease prediction.

If you tell me (1) which diseases you're targeting, (2) what data you have—EHR, labs, imaging, genomics, wearables—and (3) whether this is for research or clinical deployment, I can narrow this to 5–8 specific vendors/platforms and compare their APIs, FDA status, data requirements, deployment model, and likely fit.

When evaluating products and platforms for AI-driven disease risk identification from longitudinal patient data, the ideal choice depends on your ecosystem (provider, payer, or life sciences research), your deployment constraints, and the types of data streams you are analyzing (EHR timelines, claims, labs, or multi-omics).

Leading commercial and enterprise-grade products/platforms in this space include:

• Innovaccer Data Platform & Health AI: Known for unified patient records and advanced risk stratification models that go beyond traditional risk adjustment factors (RAF). It aggregates longitudinal clinical, financial, and social determinants of health (SDOH) data to surface rising-risk cohorts and track clinical trajectories.
• Lightbeam Health Solutions (AI Risk Stratification): Combines standard methodologies (like the Johns Hopkins ACG® System and NYU Avoidable ED) into a unified risk-scoring engine. It ingests longitudinal clinical, behavioral, and ADT (Admission, Discharge, Transfer) feeds to generate real-time alerts for patient deterioration or chronic disease escalation.
• Arcadia Analytics: A robust population health management platform featuring deep predictive analytics and patient stratification modules. It specializes in tracking polychronic patients and surfacing care gaps across multi-year longitudinal timelines.
• AyasdiCare (by SymphonyAI): Utilizes topological data analysis and advanced machine learning to find hidden patterns and high-risk patient phenotypes within massive, complex, and messy longitudinal EHR datasets without requiring manual feature engineering.
• AWS HealthOmics / Google Cloud Vertex AI for Healthcare: If you are building custom models rather than using an out-of-the-box risk stratification app, these cloud infrastructure solutions offer specialized pipelines for managing longitudinal multi-modal data, electronic health records, and genomic sequencing at scale.

To help narrow down the ideal platform, could you tell me:

• What type of organization you represent (e.g., hospital system, payer, pharma/research)?
• What primary data sources you need to integrate (e.g., EHR notes, claims data, wearable metrics, genomics)?

When evaluating artificial intelligence products and platforms to identify disease risk from longitudinal patient data (such as electronic health records, repeated lab results, and sequential clinical visits ), healthcare organizations typically look at enterprise data platforms, specialized clinical AI engines, and life sciences analytics suites. Rather than single out-of-the-box consumer apps, the market relies on scalable platforms capable of managing time-series medical data, preserving privacy, and offering predictive modeling.

Top platforms and product categories to consider include:

• Databricks Unified Data Analytics Platform : Widely utilized for ingesting real-world longitudinal data, building customized machine learning tracks via MLflow to monitor patient trajectories, and predicting specific disease or complication risks over defined time windows.
• IQVIA Orchestrated Customer Engagement & Analytics Solutions : Leverages deep healthcare-specific datasets and advanced machine learning frameworks (such as longitudinal sequence models) to assess patient risk factors and disease progression probabilities.
• AWS HealthLake / Amazon SageMaker : Provides secure, HIPAA-compliant infrastructure to store and query longitudinal patient timelines in FHIR format and apply custom or pre-built deep learning models (like recurrent neural networks) for clinical risk stratification.
• Datavant : Essential for the data-prep and de-identification phase, enabling organizations to safely connect fragmented longitudinal patient records across multiple vaults while maintaining compliance before feeding data into AI/ML risk engines.
• Google Cloud Healthcare Data Engine : Allows health systems to harmonize disparate electronic health records and leverage BigQuery ML or Vertex AI to analyze temporal clinical sequences for early identification of deteriorating conditions or chronic disease risks.

If you can share:

• The specific data modalities you have (e.g., structured EHR only vs. mixed with imaging/genomics)
• Whether you plan to build custom models or need a pre-certified clinical decision tool

I can help narrow down the ideal architectural or vendor approach for your team.

If your goal is AI that consumes longitudinal EHR data and identifies patients at elevated risk of developing disease, I’d separate the market into three buckets: turnkey clinical risk platforms, disease-specific AI, and platforms for building your own models.

Shortlist I’d evaluate

Tempus is probably the closest match to what you described. Its current platform combines longitudinal clinical records with molecular and imaging information; Tempus says it has 8.5M+ de-identified longitudinal records with linked clinical, molecular and imaging data. www.tempus.com Its provider-facing Tempus Next platform is already using longitudinal EHR, notes, imaging and time-series data to identify patients with gaps across cardiovascular diseases.

Aidoc is a different proposition: it's particularly strong when the prediction is tied to imaging or an actionable clinical workflow. Its platform includes risk stratification, EHR-fed patient management and automated identification of patients needing follow-up.

The important distinction

If you mean:

"Given 5–10 years of labs, diagnoses, medications, notes, imaging, procedures and demographics, predict which diseases this patient is likely to develop over the next 1–5 years."

then I would not start with a generic LLM or a radiology AI vendor.

I'd run a focused evaluation of:

1. Tempus — strongest commercial candidate for multimodal longitudinal prediction.
2. Epic's native predictive capabilities — especially if you're an Epic customer and want prediction embedded in workflow.
3. A custom model platform — if you have enough historical data and want your own disease-specific models.
4. Disease-specific vendors — if the target disease is something like cancer, cardiovascular disease, Alzheimer's, CKD, sepsis, etc.

There is also emerging academic evidence for general longitudinal disease forecasting: the Delphi-2M model, for example, was trained on health records from 2.3M people and reported predictions across 1,000+ diseases. But it remains a research/forecasting model rather than a turnkey clinical product.

What I'd demand in an RFP

The biggest mistake would be evaluating vendors primarily on "AI accuracy." I'd require:

• True temporal validation — train on earlier years, test on later patients.
• Calibration, not just AUROC.
• Sensitivity/specificity at clinically meaningful thresholds.
• Performance across age, sex, race/ethnicity, socioeconomic groups and sites.
• Ability to handle missing and irregularly sampled longitudinal data.
• Structured EHR plus clinical notes, medications, labs, procedures and imaging.
• Explicit handling of data leakage.
• Explainability: why did this patient get a high-risk score?
• Prospective/real-world validation rather than retrospective claims alone.
• Integration with your EHR/workflow.
• Clear FDA/clinical-device status if the output will influence diagnosis or treatment.

One particularly important point: "predict disease risk" and "identify disease" are different regulatory and clinical products. If you're going to use the output to make or materially influence a diagnosis/treatment decision, regulatory status and prospective clinical validation become central procurement criteria.

If you tell me which diseases you want to predict, what data you have (EHR/claims/labs/genomics/imaging/wearables), and whether you're a health system, payer, pharma company, or startup, I can narrow this to 5–7 vendors and compare their data inputs, models, FDA status, integrations, evidence, and likely pricing/business model.