Do multimodal clinical AI systems have enough prospective clinical evidence?

What prospective evidence actually exists for multimodal clinical AI?

Prospective evidence is strongest in two narrow areas: fracture detection on X-rays and cardiac MRI planning. In a 2023 prospective study of 1,163 exams, an AI fracture-detection tool (Gleamer BoneView) helped radiology residents find 25 additional fractures they initially missed, boosting sensitivity from 84.7% to 91.3% while keeping specificity above 97% [2]. That is a real, measured improvement in a live clinical workflow. For cardiac MRI, two prospective trials (10 healthy subjects for planning, 20 subjects including patients for shimming) found that AI-based slice planning cut total scan time by over 2 minutes (13% faster) and AI-based magnetic field shimming improved image signal-to-noise ratio by 12.5% and sharpness [3]. These are concrete, statistically significant gains in routine clinical tasks.

Outside these two areas, the prospective record is sparse. A 2025 study of an edge-AI navigation system for liver surgery achieved 99.5% precision on retrospective video data, but the authors explicitly state that 'prospective clinical trials are required to validate its clinical utility' [1]. Similarly, a multimodal AI for prostate cancer detection that combined MRI findings with clinical data (PSA, age, prostate volume) outperformed either data source alone (AUC 0.77 vs 0.70), but this was a retrospective analysis of 932 exams [6]. The pattern is clear: promising retrospective results, but very few have been tested in real-time clinical decision-making.

Why does most multimodal AI still lack prospective validation?

The main reason is that building and testing a truly multimodal system — one that fuses, say, medical images, lab results, and genetic data — is far harder than testing a single-modality AI. A 2025 scoping review of 432 multimodal AI papers found that while these models consistently outperformed single-modality ones (by an average of 6.2 percentage points in AUC), the vast majority were retrospective and used curated, complete datasets [10]. Real clinical data is messy: missing values, different formats across hospitals, and the need for cross-department coordination make prospective trials logistically daunting [8][10][11]. The same review notes that 'several challenges persist, including cross-departmental coordination, heterogeneous data characteristics, and incomplete datasets' [10].

Another barrier is that many multimodal AI systems are still in the proof-of-concept stage. For example, GPT-4V achieved 80.6% diagnostic accuracy when given both text and images from NEJM challenge cases, but the study was conducted over one week on curated public data, not in a hospital [5]. The authors themselves call for 'real-world clinical data' validation. Similarly, multimodal AI frameworks for dementia diagnosis that combine clinical data with PET scans show promise for patient stratification in trials, but the evidence so far comes from retrospective trial data, not prospective deployment [7][9]. The gap between what works in a clean dataset and what works in a busy clinic remains wide.

Where do the studies agree, and where do they conflict?

There is strong agreement across all the studies that multimodal AI — combining different data types — outperforms single-modality AI. The scoping review of 432 papers found an average 6.2-point AUC gain [10], the prostate cancer study found a 7-point AUC improvement (0.77 vs 0.70) [6], and GPT-4V's accuracy jumped from 45.2% with images alone to 80.6% with both text and images [5]. This convergence across different tasks and architectures is compelling: adding more relevant data sources consistently helps.

However, there is a clear conflict between the enthusiasm for multimodal AI's potential and the thinness of prospective proof. Review articles from 2022 and 2025 paint an optimistic picture of AI reshaping medicine [4][12], yet the actual prospective studies here are limited to fracture detection [2] and cardiac MRI [3]. The liver surgery navigation system [1] and the dementia stratification tools [7][9] are still awaiting prospective trials. This is not a contradiction — it is a field that is moving fast on the research side but has not yet delivered on the deployment side. The honest takeaway is that prospective evidence exists for a few specific, well-defined tasks, but for the broader vision of multimodal AI in medicine, it remains largely unproven in real clinical settings.

About These Sources

This answer is built on 12 peer-reviewed studies — published from 2022 to 2026, 7 from 2024 or later, 9 in Q1 journals, collectively cited 3,273 times — selected as the most relevant from 13 studies that passed quality screening, drawn from 51 papers retrieved from a database of over 500 million.

Sources used in this answer

Related research questions