Machine Learning Research Fellow - Dermatology Equity AI (ISIC Archive Data)

We're looking for a researcher to produce an original, TRIPOD+AI-compliant finding on skin-tone-fair melanoma and keratinocyte-cancer triage, using a large public dermoscopic image archive, cross-validated against datasets specifically curated for skin-tone diversity. This is an invited submission to JPMedAI (Journal of Precision Medicine and Artificial Intelligence), a new, open-access, peer-reviewed journal currently assembling its inaugural issue. Article processing charges are waived during this launch phase, meaning there is no cost to you to publish. We're saying that plainly because we want the right fit: this suits researchers who value a genuine, citable publication and hands-on methodological support as much as (or more than) the cash fee, which reflects that context. THE DATASET Training dataset: the ISIC Archive, fully open, no registration, over 69,000 publicly released dermoscopic images. External validation: datasets specifically curated for skin-tone diversity and biopsy-confirmed diagnoses, spanning the full Fitzpatrick range, plus an additional Brazilian cohort with skin-tone metadata, all fully open. The angle: dermatology-fairness is already an active research area, so we're not claiming virgin territory here. The opening is in reporting depth: most existing fairness work in this space reports an overall AUC or an AUC "gap" between skin tones. We have a specific direction in mind involving calibration and decision-curve analysis reported separately by skin type, plus an audit of what the model is actually attending to across different skin tones, but haven't fixed it here. Propose your own take in your application. KEY CONSIDERATIONS Restrict the primary malignant/benign endpoint to images with confirmed histopathology. At least one of the diversity-focused datasets has a well-documented label-noise problem if used as ground truth rather than as a fairness-focused test set. The differentiator here is methodological rigor (subgroup calibration, decision-curve analysis, an artifact audit for confounders like rulers or ink markings across skin tones), not the topic itself. Say this explicitly in the introduction so the contribution is clear against the existing fairness literature. POST-SUBMISSION PEER-REVIEW REVISIONS Following submission, JPMedAI's peer review may request revisions before acceptance. Standard revisions (clarifications, additional citations, supplementary analyses within the existing study design) are expected as a normal part of authorship and are not separately compensated, consistent with standard academic publishing practice. If reviewers request substantially new work (e.g., a new validation dataset or a materially different analysis), a further milestone will be added at that time, sized to the actual scope requested. Response to reviewers should come from you as the named author; we'll support the process editorially throughout. GENERAL EXPECTATIONS FOR THE RESEARCH FELLOW Use a large, freely or quick-access public dataset for model development (training) and at least one genuinely independent dataset for external validation. Report according to TRIPOD+AI (2024), the AI/ML-specific extension of TRIPOD, citing the original TRIPOD statement as the foundational reference. Grad-CAM should be used for convolutional-image attention-map / saliency-based explainability. Report discrimination (AUC), calibration (plots, not just calibration slope/intercept in text), and decision-curve analysis against a clinically meaningful comparator, not against "no model" alone. Explicitly test and report subgroup performance by Fitzpatrick skin type rather than a single pooled metric. Before finalising the research question, complete a focused, dated literature search on the exact proposed endpoint/model combination and document the closest 2–3 prior papers and the specific point of difference. Write up the work as a scientific manuscript, 3,000–6,000 words, up to 8 figures, in this structure: Abstract (≤300 words, structured), Keywords (3–6), Introduction, Methods, Results, Discussion, Conclusion, References (Vancouver style). JPMedAI provides free academic writing support to any author who needs it, separate from the project-specific editorial guidance described below. WHAT YOU'LL GET An invited submission pathway to JPMedAI, a new, open-access, peer-reviewed journal building its inaugural issue. Early submissions get direct editorial engagement, not a slot in a large backlog. Hands-on editorial and methodological support, beyond the journal's own free academic-writing service noted above, we'll personally guide you on statistical reporting, clinical framing, and endpoint interpretation to bring your first draft to publishable, TRIPOD+AI-compliant standard. Named byline, full citation credit, ORCID-compatible publication metadata, and a PDF for your portfolio, CV, or dissertation appendix. A $100 honorarium on completion/acceptance. The primary value on offer here is the invited, mentored publication itself. First-mover advantage: be among the first researchers to publish on this specific dataset and question. WHAT THIS IS NOT This is not content writing. You are not ghost-writing. You are an independent researcher producing original work under your own name. The peer-review and editorial process exists to enforce scientific governance standards, not to change your conclusions or direct your inquiry. IDEAL CANDIDATE PROFILE We expect interest from two kinds of applicants, and both are genuinely welcome: Early-career researchers (PhD students, postdocs, recent graduates) who already have at least one preprint or publication and want a fast, well-supported second one. Strong ML / data science practitioners without a formal clinical or health-publication track record who want to break into applied health research. For this group, our editorial support specifically includes guidance on clinical framing, endpoint selection, and TRIPOD+AI reporting norms. Also useful (but not mandatory): background in dermatology, health equity research, or computer vision; experience with CNN-based image classification; familiarity with calibration plots, decision-curve analysis, or explainability methods. Sharing workload with one or two collaborators who will be coauthors is welcome. IF YOU ARE INTERESTED Please answer the questions below.