Patrick Iben, MS · Applied mathematics, research engineering, predictive analytics, and clinical biostatistics
I work to eliminate needless ambiguity, risk, and suffering in decisions that rest on data. I build mathematical frameworks and turn uncertain signals into inspectable, cross-validated systems across clinical trial design and analysis, causal and network inference, model and AI evaluation, and decision support built on public and sensor data. I look for agreement across independent lines of evidence, and for the shared structure that connects fields that seem unrelated, so a method proven in one setting can be tested honestly in another.
7 years of clinical biostatistics in contract research (early-phase clinical pharmacology, CDISC/ADaM, PK/NCA) and 12 years in research. Remote US.
Contact: patrickiben@gmail.com · LinkedIn · ORCID 0009-0005-2443-7973
Resume (2-page PDF) · Full CV (4-page PDF) · Word versions
Research engineering and systems architecture: translating methods across domains, integrating data and models, and building tools whose assumptions and evidence can be inspected.
-
Coastal Water Resilience (source): Connects my CRT framework and network-contagion research to infrastructure planning. Real public station records, terrain, evacuation zones, Census context, and NOAA/NWS observations and forecasts support projected-impact screens and ten-parameter response plans. Actions explain why each parameter matters at that phase, with timing and escalation steps. Unsupported dimensions stay explicit; the synthetic exercise is separately labeled.
-
Public Data Consoles (source): 36 browser dashboards spanning hazards, public health, access to care, and markets, including NHS Scotland Strain, U.S. Grid Strain, River Flood Propagation, Community Resilience, and three place finders: Food Desert Oases, Free Clinic Finder, and Free Public Bathroom Finder. Visitors need no API key. Public feeds and scheduled data publication support dated, source-aware displays; all 36 include ten-parameter response plans with priorities, escalation, and plain-language explanations for public users and professional operators.
-
Autonomic Telemetry (source): Wearable sensing from Bluetooth devices through an in-memory Mac bridge to a mobile browser. Work/Rest views combine a personal reference with explicit signal quality, freshness, and source coverage. The Signal Lab adds replayable numerical views with analytic and regression checks. Live wearable input runs through the local bridge.
-
Resilience Telemetry (research guide): An application of my directed-contagion research to cross-asset volatility networks: spillover analysis, transmitter stability, historical replay, and held-out comparisons against loudness and persistence baselines. Public-data ingestion and the documented method correspondence make the adaptation inspectable.
-
myco-fusion (source): How far a phone camera can go toward a reflectance spectrum: a local research workbench and a reproducible simulation study on measured spectra. On 532 natural-object samples, a simulated camera met both draft targets for 83% when calibrated on similar objects and 60% with a color chart; measured curves from six iPhones gave similar results. No phone photo or wild fungus was tested, and it is not for judging whether anything is safe to eat.
Clinical trial tools: worked examples, templates, apps and checks for biostatistics consultants and employees, from sample size to CSR text.
-
Trial biostatistics toolkit (source): open-source tools, templates and worked examples across study planning, programming, operations and reporting, focused on early-phase clinical pharmacology. All 23 modules open from a browser catalog; core study setup and analyses need no local R, Python, account or download. Each page states what was executed (runs, parse-checked, not executed); example data are synthetic, and nothing is validated or regulatory.
- Flagship worked example: one synthetic 2x2 crossover bioequivalence study (N = 40), from sample size and randomization through the SAP, ADaM (admiral), NCA (PKNCA), the ICH M13A analysis, TLFs, and CSR Section 11-style text, regenerated by one script in CI. QC is a same-author re-derivation, not independent double programming.
- TRIALMON safety review: a Safety Review Committee packet and Shiny console for a synthetic first-in-human dose escalation, with a Hy's Law grid that checks whether each flag holds under alternative SAP conventions. Not validated; author sign-off pending.
- QC harness: ten gates, including a same-author recomputation of the PK, bioequivalence, and adverse-event results and a seeded-defect check of 18 defects registered in advance (counts for that list, not a detection rate), in which only the recomputation caught the seven numeric and counting defects. The ledger records the eight automated checks as passing and the two human-review gates (QC review and AI-use accountability) as reviewed; the checks cover the toolkit's own libraries and synthetic data, not any study.
- Method transfers from my research: an exploratory residual-carryover study of the ICH M13A 5% pre-dose rule, following the design of my CRT simulation study; an exploratory C-QTc hysteresis check, a small formalization of the standard check on the established signed loop area, framed from my directed-contagion work; and the seeded-defect check (established error seeding, adapted from the teeth check in my consilience engine) and the Hy's Law grid (a structured sensitivity analysis, adapted from its method-invariance grid).
- SAS and R TLF libraries (221 programs per language across five study designs; all 221 R programs execute on synthetic fixtures, native SAS not executed) and four R Shiny apps that run in the browser, with live loading status and a pre-deploy browser test: TRIALMON console, IRT/RTSM UAT tracker, Study Lifecycle Monitor, and the vaccine analysis workspace, which maps study CSVs or loads a fictional example and runs six immunogenicity and reactogenicity outputs with no installation or account. The operational risk signals and hours reconciliation apps also run in the browser with mapped study inputs.
- eCRF review workspace (source): four browser tools for your own CSV extracts: coding QC assist, de-identification review assist, participant risk screening, and anomaly and drift screening. Each tool starts blank; no client data or licensed MedDRA dictionary is bundled, and the optional example is synthetic. The evaluation documents the original implementation's measured failures (the de-identification stand-in left 108 of 308 identifiers readable, the full model 91) and a pre-registered within-sibling check, adapted from my AlphaGenome audit, that routes both term families to human coding. Not validated. The apps moved here from the archived ecrf-pipeline-apps repository, whose pages now redirect.
- Point-and-click setup: an optional local control panel for private projects, numerical examples, QC checks and generated study packages, with guided forms that map CSV inputs to the TLF libraries and monitoring checks.
- Operational tools, no AI required: Open Study Tracker, a coverage and continuity toolkit, and a find-hours reconciliation for effort only.
-
catmech: An R package (version 1.0.1) for regression when recorded values may be wrong and an audit has checked a subset against the source. Its own tests on public data found that simple linear multiple imputation given the recorded value matched or beat the recording-error model the package added, so linear imputation is the recommended start. The clustered-audit option is experimental, and the audit-planning function failed a rule fixed in advance. Not on CRAN, with no outside peer review or clinical validation.
-
calibrated-clinical-ml: From-scratch calibration and decision-curve evaluation against a synthetic oracle; demonstrates how identical AUC can conceal different probability calibration.
Signal instruments and experiments: browser tools that read a body, sound, or motion signal and label what is measured, inferred, or simulated.
-
Body Signal Labs (source): Browser scenes that follow signals from the body and say what was measured and what was inferred. Endless Cockpit sets riding speed through nine procedural scenes from pedaling cadence (a Bluetooth sensor, a smart trainer, or an on-screen pedal) and exports sessions as TCX activity files. Biofeedback Path Cockpit adds heart-rate and Muse S EEG inputs, a paced-breathing recovery view, and a heart-rate-variability coherence game, and uses Autonomic Telemetry's engine for its heart-shift score. Sensors report their connection state, and demos and replays are labeled. No page is a medical device, and no real cadence sensor or heart-rate strap has connected to either cockpit yet.
-
Generative AV Labs (source): Browser instruments for light and sound that run with no helper app or server; audio, MIDI, and pictures stay in the browser. Nebula draws an electronic drum kit's audio and MIDI as a twelve-second spiral and matches a picture to a sound by sliced optimal transport. Displacer blends two sounds by moving energy along frequency with an optimal-transport map instead of a crossfade. Chimera Synth grows a bass voice from a 20-gene genome. Enter turns a picture into a depth relief to walk through and a sound into a map to fly through; it uses a depth-estimation model, downloaded only on request, and the other four use no model. Detail Forge Analog adds a synthesized high-frequency layer that is procedural, not recovered audio. No real drum kit has been used with Nebula yet.
-
PetRand (motion game): An experimental entropy source designed around a motorized cat-toy ball bouncing in a small arena, with the tools to test whether it produces unpredictable numbers: two health tests that follow the designs in NIST SP 800-90B, an entropy credit per sample, and HMAC-SHA-256 mixing into a 32-byte seed. Research prototype: no physical arena has been built or measured, runs use a labeled simulator unless a camera or microphone is selected, the entropy credit is a placeholder, and nothing is certified.
Manuscript → code → demo map: manuscript versions, research artifacts, teaching materials, and related applications for seven studies. Each research project has a concise study guide and a complete educational companion, linked below as public PDFs. The AlphaGenome entry distinguishes the current rework from the earlier audit. Its status was updated on 29 October 2026, and the frozen-rules entry was added on 11 November 2026; the other submission and review statuses retain their 23 October 2026 record.
-
Contamination-Resilient Cluster Randomized Trial Design under Spatial Interference: A Simulation Study of Contamination Control. Statistics in Medicine (submitted 19 October 2026). Latest preprint · Code and data · Live demo · Study guide · Complete companion
-
Auditing preprocessing-sensitive allele rankings in predicted chromatin contact maps. BMC Bioinformatics (submitted 29 October 2026; technical check). Latest manuscript · Code and data · Live demo · Study guide · Complete companion
Earlier audit, desk-rejected for journal fit: What do within-site comparisons of AlphaGenome Atlas predictions measure? An audit at ClinVar sites in 11 genes. Earlier preprint · Archived code and data · Earlier demo · Earlier teaching materials.
-
A false floor: Refuting our own information-limit claim in a lethal identification task. PLOS ONE (submitted 22 October 2026). Latest preprint · Provenance and training-catalog overlap records · Earlier code snapshots · Study guide · Complete companion
-
Adaptive decision support under frozen rules: guarantees, failure modes, and component admission. PLOS ONE (submitted 11 November 2026). Latest preprint · Replication data · Fine-tuned checkpoint · Study guide · Complete companion
-
Anticipating Network Contagions: When Directedness Outperforms Connectedness. Applied Network Science (with editor). Author draft · Code and data · Resilience Telemetry · Study guide · Complete companion
-
Which data modality recognizes a gene set? A domain-agnostic consilience engine. PeerJ (in peer review). Author draft · Code · Study guide · Complete companion
-
No emotion-specific image-computable code in the human left amygdala. Peer Community In, preprint review. Latest preprint · Code · Study guide · Complete companion
Published work
-
Acknowledged contributor: developed the data extraction and normalization programs for the study dataset in Osterhage KP, Hser YI, Mooney LJ, et al. Identifying patients with opioid use disorder using International Classification of Diseases (ICD) codes: Challenges and opportunities. Addiction. 2024;119(1):160-168.
-
Co-author: Ben-Avie M, Rosenwald L, Depastino R, Iben P. How can we help young adults with autism thrive in the workplace? The Conversation, 2016.
Small, runnable tools for examining model-output reliability and service behavior:
-
numguard: Deterministic checks of extracted numeric values against allowed facts, required facts, and a finite phrase denylist; JSON/JUnit export for CI.
-
tool-grounded-agent: A scripted planner demonstrates tool-ledger provenance and a grounding backstop. The harness includes executable known-gap tests; it does not run a language model.
-
grounded-rag: Lexical retrieval and an extractive answerer with checks for numeric membership and lexical support. A runnable evaluation prototype.
-
guarded-llm-service: The grounding guard served over FastAPI with Docker, CI, and Prometheus-style metrics.
Shared examples and hosted demonstrations use public or synthetic data; the vaccine workspace, the eCRF review workspace, and the other toolkit browser apps also accept mapped study CSVs, computed in the browser. Clinical examples are synthetic and labeled by status; none is validated or regulatory, and the in-browser clinical apps have been checked in Chrome only. Consoles distinguish observations, source-issued forecasts, modeled estimates, and exercises, with freshness and coverage stated. Autonomic Telemetry also supports private local wearable input; research and prototype limits are documented. The signal instruments label demo, replay, and simulated input.


