Free scoping read. A written assessment of your manuscript in two business days, at no cost.

Start one
Contents

Academy · Research Data, Analytics & AI · Deep Learning in Healthcare

Deep Learning for Predictive Analytics in Healthcare: Evidence from Two Landmark Studies

Two of the most frequently cited studies in clinical deep learning, a dermatologist-level skin cancer classifier and a hospital-scale EHR mortality predictor, demonstrate what these models can verifiably achieve and what their reported figures signify.

Pubrica AcademyWritten by a healthcare data science research writer Reviewed bya clinical informatics researcher Published10 October 2026 Last checked10 October 2026 Reading time7 minutes Sources10 peer-reviewed

“Deep learning for healthcare” is used sufficiently loosely in marketing material that the claim should be anchored to specific, verifiable papers. Two studies are frequently cited because their methods and results are fully reported and independently reproducible in substance: the 2017 Nature paper by Esteva and colleagues on convolutional neural network (CNN) classification of skin lesions, and the 2018 npj Digital Medicine paper by Rajkomar and colleagues applying deep learning directly to raw electronic health record (EHR) data across two U.S. academic medical centers. Together they illustrate the two dominant architectural families in healthcare predictive analytics, namely CNNs for image-based diagnosis and recurrent or sequence models for longitudinal patient data, and provide reported accuracy figures rather than illustrative ones.

Learning objectives

  • The training and testing of a CNN (GoogLeNet Inception v3) against board-certified dermatologists in skin cancer classification, with the AUC figures reported.
  • The performance of deep learning applied to raw, FHIR-formatted EHR data on four distinct clinical prediction tasks across two hospital systems.
  • The reasons both studies report performance as AUC/AUROC against specific test sets, and the implications and limitations of this for real-world deployment.

Common misconceptions

  • “Deep learning models in these studies universally 'outperform doctors'.” Esteva et al. reported performance on par with board-certified dermatologists on the specific binary classification tasks tested (keratinocyte carcinoma vs. benign seborrheic keratosis, and malignant melanoma vs. benign nevi) under controlled test conditions, not superiority in general clinical practice or across the full range of dermatologic diagnosis.
  • “A single AUC value indicates how the model performs in a live hospital.” Rajkomar et al. reported AUROCs of 0.93-0.94 for in-hospital mortality prediction across two separate medical centers using retrospective data from 216,221 adult hospitalizations. This is a strong result on held-out historical data, but the paper itself frames it as a predictive-modeling benchmark rather than a validated clinical deployment outcome.

Convolutional neural networks (CNNs) for image-based diagnosis: the skin cancer study of Esteva et al.

Esteva A, Kuprel B, Novoa RA, Ko J, Swetter SM, Blau HM and Thrun S published "Dermatologist-level classification of skin cancer with deep neural networks" in Nature in 20171. The team used a GoogLeNet Inception v3 convolutional neural network, pretrained on roughly 1.28 million general images (ImageNet) and then fine-tuned via transfer learning on a dermatology-specific dataset of 129,450 clinical images spanning 2,032 distinct skin diseases, organized into 757 training classes1.

The model was evaluated against at least 21 board-certified dermatologists on biopsy-confirmed test images. On the task of distinguishing keratinocyte carcinomas from benign seborrheic keratoses, the model achieved an AUC of 0.96 on a 707-image test set. On distinguishing malignant melanoma from benign nevi, it likewise achieved an AUC of 0.96 on a 225-image test set, with a dermoscopy-image comparison on a separate 1,010-image set yielding an AUC of 0.941. In both binary tasks, the CNN’s sensitivity-specificity performance matched the dermatologists’ point estimates on the same images, which is the specific, narrow claim underlying the “dermatologist-level” result.

Sequence models applied to raw electronic health record (EHR) data: the study of Rajkomar et al.

Rajkomar A, Oren E, Chen K and colleagues published "Scalable and accurate deep learning for electronic health records" in npj Digital Medicine in 20182. Rather than hand-engineering features from structured EHR fields, the team represented each patient’s entire record, including clinical notes, using the Fast Healthcare Interoperability Resources (FHIR) standard and applied deep learning models directly to that raw representation, drawing on 216,221 adult patients hospitalized for at least 24 hours across two U.S. academic medical centers. The resulting dataset, once unrolled into individual data points, totalled over 46 billion data points2.

The models were evaluated on four prediction tasks, each measured as AUROC, and compared against traditional predictive models built on curated features. Reported results across the two sites were: in-hospital mortality, 0.93-0.94; 30-day unplanned readmission, 0.75-0.76; prolonged length of stay, 0.85-0.86; and final discharge diagnoses, 0.90 (frequency-weighted)2. The paper reports that the deep learning approach outperformed the traditional baseline models on every task tested, while also requiring substantially less manual feature engineering because it worked from the raw record rather than pre-selected variables.

Correspondence between model architecture and data type

The two studies are instructive as a pair because the choice of architecture follows from the structure of the data rather than from fashion. Esteva et al.’s task is a fixed-size image classification problem, for which convolutional architectures are designed: stacked convolutions learn spatially local visual features (edges, textures, lesion borders) that combine into higher-level patterns. Rajkomar et al.’s task is a longitudinal, irregularly timed sequence of clinical events for each patient, which is why their approach centers on sequence-capable deep learning models applied to a time-ordered FHIR representation rather than a single static feature vector.

This distinction is important when scoping a predictive analytics project in healthcare: a CNN pipeline designed for radiology or dermatology images will not transfer to a readmission-risk problem based on structured and free-text EHR data, and vice versa. These two papers are appropriate examples when a stakeholder asks whether such methods have been shown to work, because both report test-set performance against meaningful comparators rather than internal validation alone.

Interpretation of AUC/AUROC claims without overstatement

Both papers report AUC (area under the receiver operating characteristic curve) as their primary performance metric, which summarizes a model’s ability to rank positive cases above negative ones across all possible decision thresholds. An AUC of 0.96, as in the Esteva skin-cancer comparisons, or 0.93-0.94, as in the Rajkomar mortality predictions, is a strong discriminative result on the specific test set used, but AUC does not by itself describe calibration, the effects of real-world prevalence or performance drift once a model is deployed outside the population on which it was trained and tested. Authors citing either paper in a grant application, systematic review or internal evaluation should report the specific task, test-set size and comparator exactly as the original authors framed them, rather than compressing “AUC 0.96 on a 707-image biopsy-confirmed test set” into an unqualified claim that the model “beats doctors.”

Reporting requirements of TRIPOD+AI for clinical prediction models

Both Esteva et al. and Rajkomar et al. predate the specific reporting guideline now used to standardize how studies like theirs should be written up. Collins, Moons, Dhiman and colleagues published the TRIPOD+AI statement in BMJ in 2024, updating the original 2015 TRIPOD statement to explicitly cover prediction models built with regression or machine learning methods, including deep learning.3 TRIPOD+AI's checklist asks authors to report, among other items, how the model's outcome and predictors were defined and measured, the exact sample size and number of events relative to candidate predictors, how missing data were handled, and — specifically relevant to both headline studies above — the model's discrimination and calibration performance, not discrimination (AUC/AUROC) alone.3 That calibration requirement is the gap this article flagged earlier: an AUC of 0.96 or 0.93-0.94 describes ranking ability, and TRIPOD+AI's own framework is explicit that discrimination has to be reported alongside calibration for a reader to judge whether the model's predicted probabilities are themselves trustworthy, not just well-ordered.

Continuing reporting obligations for deployed models after publication

TRIPOD+AI's predecessor, the original TRIPOD statement published by Collins, Reitsma, Altman and Moons in Annals of Internal Medicine in 2015, was built around a single development-and-validation study design.4 TRIPOD+AI extends that scope specifically because machine learning and deep learning models, unlike the regression models TRIPOD 2015 was written for, are commonly retrained or updated after initial deployment — which is exactly the scenario neither Esteva et al.'s nor Rajkomar et al.'s original papers needed to address, since both reported a single retrospective evaluation rather than a model already live in clinical use.3 For any deep learning system that moves from a published benchmark like these two toward actual deployment, TRIPOD+AI's expanded reporting expectations — and not just the original paper's AUC figure — become the relevant standard for what has to be disclosed about how the model was built, validated, and is intended to be monitored and updated over time.

That distinction — a model evaluated once on a fixed historical dataset versus a model expected to keep learning or be retrained after it ships — is also the reason citing Esteva et al. or Rajkomar et al. as proof a current, commercially deployed product works is a category error even when the underlying architecture is similar. Neither paper reports on a model under an active change-control process; both report a single, dated snapshot of performance. A literature review, grant narrative, or internal evaluation that wants to make a claim about a live deployed system needs a source that actually describes that system's post-deployment monitoring and update plan, not a citation to either landmark benchmark by proxy.

Regulatory and data-infrastructure context of deployment

Both studies this article has discussed were research publications, not deployed products, and the gap between the two is largely a regulatory and data-infrastructure one. FDA's Digital Health Center of Excellence sets out the agency's current approach to artificial intelligence and machine learning in software as a medical device, including the predetermined change control plan mechanism that lets a manufacturer pre-specify how a model may be updated after clearance without triggering a brand-new submission for every retraining cycle.5 That regulatory question only arises once a model is trained on usable data in the first place, which is where data quality becomes the binding constraint. Weiskopf and Weng's framework for electronic health record data quality assessment — completeness, correctness, concordance, plausibility, and currency — describes exactly the dimensions a deep learning pipeline trained on EHR data inherits uncritically unless they are checked.6 Cowie and colleagues' review of how electronic health records facilitate clinical research documents the same interoperability and governance issues that determine whether a multi-site training dataset is usable at all.7

Where the input to a model is unstructured clinical text rather than structured fields, Wieland-Jorna and colleagues' systematic review of natural language processing systems for extracting information from electronic health records found extraction performance varying widely by clinical domain, a caution against assuming an NLP preprocessing step feeding a downstream deep learning model is uniformly reliable.8 And on the scale of adoption behind these technical questions, IQVIA Institute market-growth data on digital health tools documents the pace at which AI-assisted diagnostic and predictive tools are expanding in scope and clinical use, the backdrop against which the data-quality and reporting standards discussed in this article matter at a population, not just single-study, scale.9 Finally, on the data-capture side that ultimately feeds both research datasets and any later deployed model, Harris and colleagues' description of REDCap documents a metadata-driven electronic data capture platform used widely to assemble exactly the structured, auditable training and validation datasets that predictive analytics work of this kind depends on.10

Key points

Key points

Two well-documented studies, a CNN matching dermatologists on specific skin-lesion classification tasks (AUC 0.96) and a sequence-based deep learning model predicting hospital outcomes directly from raw EHR data (AUROC 0.93-0.94 for mortality), demonstrate genuine, reproducible deep learning performance in healthcare, provided that the claims are reported with the same task-specific precision used in the original papers.

Practical application

Practical application

  1. The architecture should be matched to the data structure before a precedent is cited. Image-based diagnostic tasks draw on CNN precedents such as Esteva et al., whereas longitudinal patient-record tasks draw on sequence-model precedents such as Rajkomar et al.; citing an inappropriate precedent weakens a methods section.
  2. AUC/AUROC should be reported with its exact test-set context. The test-set size, the comparator (for example, board-certified dermatologists or traditional feature-based models) and the specific task should be stated, rather than a bare percentage.
  3. Retrospective benchmark performance should be distinguished from deployment validation. Both landmark studies are retrospective evaluations on held-out historical data; neither is a prospective clinical deployment trial, and a literature review should state this explicitly.

Frequently asked questions

No. It demonstrated that a CNN could match board-certified dermatologists' performance on two specific binary classification tasks (keratinocyte carcinoma vs. benign seborrheic keratosis, and melanoma vs. benign nevi) using biopsy-confirmed test images — a narrower and more specific claim than general diagnostic replacement.

It applied deep learning directly to a FHIR-based representation of the raw patient record, including clinical notes, rather than requiring researchers to hand-engineer a smaller set of structured features first, and it reported gains over traditional feature-based baselines on all four tasks tested.

Yes, for prediction models using regression or machine learning methods: TRIPOD+AI is the current guidance, and the 2015 TRIPOD statement and its explanation and elaboration document remain the reference for the original regression-only scope and for historical studies reported under it.

AUC/AUROC measures ranking ability, that is, whether higher-risk cases are scored above lower-risk ones, but it does not indicate whether the predicted probabilities themselves are accurate; TRIPOD+AI requires both discrimination and calibration to be reported because a model that ranks well may still be poorly calibrated.

References

  1. Esteva A, Kuprel B, Novoa RA, Ko J, Swetter SM, Blau HM, Thrun S. Dermatologist-level classification of skin cancer with deep neural networks. Nature. 2017;542:115-118. read 10 October 2026. View on Nature
  2. Rajkomar A, Oren E, Chen K, et al. Scalable and accurate deep learning with electronic health records. npj Digital Medicine. 2018;1:18. read 10 October 2026. View on npj Digital Medicine
  3. Collins GS, Moons KGM, Dhiman P, et al. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ. 2024;385:e078378. read 10 October 2026. View on EQUATOR Network
  4. Collins GS, Reitsma JB, Altman DG, Moons KG. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the TRIPOD statement. Ann Intern Med. 2015;162(1):55-63. read 10 October 2026. View on EQUATOR Network
  5. U.S. Food and Drug Administration. Digital Health Center of Excellence: Artificial Intelligence and Machine Learning in Software as a Medical Device. read 10 October 2026. View on FDA.gov
  6. Weiskopf NG, Weng C. Methods and dimensions of electronic health record data quality assessment: enabling reuse for clinical research. Journal of the American Medical Informatics Association. 2013;20(1):144-151. read 10 October 2026. View on PubMed Central
  7. Cowie MR, Blomster JI, Curtis LH, et al. Electronic health records to facilitate clinical research. Clinical Research in Cardiology. 2017;106(1):1-9. read 10 October 2026. View on PubMed Central
  8. Wieland-Jorna Y, van Kooten D, Verheij RA, de Man Y, Francke AL, Oosterveld-Vlug MG. Natural language processing systems for extracting information from electronic health records about activities of daily living: a systematic review. JAMIA Open. 2024;7(2):ooae044. read 10 October 2026. View source
  9. IQVIA Institute for Human Data Science. Digital Health Trends: market growth in digital health applications. read 10 October 2026. View source
  10. Harris PA, Taylor R, Thielke R, Payne J, Gonzalez N, Conde JG. Research electronic data capture (REDCap): a metadata-driven methodology and workflow process for providing translational research informatics support. J Biomed Inform. 2009;42(2):377-381. read 10 October 2026. View on PubMed

Every source above is a peer-reviewed article, a publisher’s record of one, or a standard issued by the body that owns it. Each carries the date it was last read. Where a standard moves, this article is revised rather than left up with a note.

How to cite this article

Pick your referencing style and copy it. Your university or target journal may vary the punctuation — check its guide if it has one.

Pubrica Academy. (2026, October 10). Deep Learning for Predictive Analytics in Healthcare: Evidence from Two Landmark Studies. Pubrica. https://pubrica.com/academy/predictive-analysis/deep-learning-for-predictive-analytics-in-healthcare/

This article is written to be learned from. Cite it if you are referring to what it says; if you are using the evidence behind it, cite the sources listed above instead.

Checklist for citing clinical deep-learning results

This checklist should be used before an AUC or accuracy figure from a healthcare deep-learning paper is quoted.

[ ] Confirmed the architecture (CNN, RNN/sequence model, etc.) matches the data type you are discussing
[ ] Quoted the exact test-set size and composition alongside the AUC/AUROC figure
[ ] Named the specific comparator (clinicians, traditional model) used in the original study
[ ] Distinguished retrospective benchmark results from prospective clinical validation
[ ] Checked whether calibration, not just discrimination (AUC), was reported
[ ] Confirmed whether TRIPOD+AI (ML models) or the 2015 TRIPOD statement (regression) is the applicable reporting standard
[ ] Linked back to the primary paper rather than a secondary summary

Pubrica AcademyWritten and maintained by Pubrica’s subject matter experts. Articles are revised when the underlying standards change, and each article states the date on which it was last reviewed.

How Pubrica can help

Pubrica’s data science writers trace every deep learning performance claim to its primary source and report it with the exact task, test set and comparator framing of the original study.

Find something else

Search the Pubrica Academy

Access is free and requires no account. Enter a topic, or open the full library to filter by category, sub-topic or resource type.

Grounding a predictive analytics literature review in verifiable deep learning evidence

A scoping assessment is free and takes two business days. It provides a written assessment of what the work requires and which specialist would undertake it.

Get a free quote

Regulations

Check your project against this rule

  1. 1Your work
  2. 2Your details
  3. 3Send
What do you need?

Material you send is held under signed confidentiality, with access limited to the assigned team and encrypted transfer. It is not used to train anything and is not passed to third-party tools. Ask for the agreement before you send anything you would rather cover first.