Kurzweil Scorecard: Proteomic Patterns and Diagnostic AI in 2003

In Chapter 9 of The Singularity Is Near — the “Response to Critics” chapter, where Kurzweil answers skeptics who say his forecasts outrun reality — he reaches for two concrete examples of AI already in the clinic. Both examples come from 2002–2003. Both were supposed to prove the rest of his book was not science fiction. Twenty-three years later, one of them is a cautionary tale about statistical hygiene, and the other is a quiet, continent-sized build-out that now encompasses more than a thousand FDA authorizations. The thesis underneath both was right. The specific objects of the thesis did not survive.

The predictions

Kurzweil’s two claims in this batch are retrospective — assertions about what had already happened by the time he wrote the book:

“Serum proteomic patterns in blood had been shown by 2002 to identify prostate and ovarian cancer.” (ch. “Response to Critics”)

“Commercial AI-assisted medical diagnostic products such as TriPath’s FocalPoint existed by 2003.” (ch. “Response to Critics”)

The first refers to a February 2002 Lancet paper by Emanuel Petricoin, Lance Liotta, and colleagues at the NCI-FDA Clinical Proteomics Program — a study that used mass-spectrometry pattern recognition on serum and reported, in their validation set, correctly calling 50 of 50 ovarian cancers and 63 of 66 controls. The second refers to Neopath’s AutoPap system, which the FDA cleared in 1998, and which TriPath Imaging inherited and marketed as FocalPoint — an automated cervical-cytology pre-screener that ranked Pap slides by suspiciousness so cytotechnologists could focus on the top decile.

In The Singularity Is Nearer (2024), Kurzweil returns to the thread. He writes that “in 2021 a Johns Hopkins team developed an AI system called DELFI that is able to recognize subtle patterns of DNA fragments in a person’s blood to detect 94 percent of lung cancers via a simple lab test — something even expert humans cannot do alone.” That sentence is the 2024 echo of the 2002 proteomics claim. A few pages later, on diagnostic AI broadly: “As the 2020s progress, AI-powered tools will reach superhuman performance levels at virtually all diagnostic tasks.”

Where we actually are

Prediction 1: The proteomic pattern test that ate its own tail

The Petricoin Lancet paper is, by one measure, wildly successful: it has been cited 3,249 times, the most-cited paper in our literature index on serum proteomic cancer detection from 2002 to 2010. By another measure, it is one of the most thoroughly refuted results of the 2000s proteomics era.

Within three years, Keith Baggerly and Kevin Coombes at MD Anderson reverse-engineered the raw data and published “Signal in Noise: Evaluating Reported Reproducibility of Serum Proteomic Tests for Ovarian Cancer” (JNCI, 2005, 296 citations in our index). Their audit found that cancer and control samples had not been run in randomized order on the mass spectrometer — all controls went first, all cancers later — and a mid-run instrument drift correlated with the signal the classifier was picking up. In plain English: the test was partly learning whether a sample had been run before or after the machine needed recalibration.

Correlogic Systems, the company formed around the method, had an exclusive agreement with LabCorp to commercialize the test as OvaCheck with a planned April 2004 launch. The FDA intervened, arguing that the test met the definition of a medical device and required clearance. Correlogic filed for Chapter 11. OvaCheck never reached patients.

What is striking is that the direction of Kurzweil’s prediction survived while the specific exemplar did not. Blood-based, pattern-recognition cancer detection is exactly the enterprise that now runs in production — just not via SELDI-TOF protein peaks. The replacement mechanism is cell-free DNA.

The most-cited recent paper in our index on this theme is Mathios et al., “Detection and characterization of lung cancer using cell-free DNA fragmentomes” (Nature Communications, 2021, DOI: 10.1038/s41467-021-24994-w, 430 citations). It reports the DELFI result Kurzweil highlights in the 2024 book: 94% cancer detection across stages at 80% specificity, including 91% sensitivity on stage I/II disease, using whole-genome sequencing of plasma cell-free DNA and a machine-learning classifier trained on fragment-length distributions. Liver and early-stage lung variants of the same approach follow — Foda et al. 2022 for liver (141 citations), Yang et al. 2021 for primary liver cancer (103 citations).

On the commercial side, GRAIL’s Galleri test reported top-line PATHFINDER 2 results in October 2025: 73.7% episode sensitivity across the twelve cancers responsible for two-thirds of US cancer deaths, 40.4% across all cancers, and 99.6% specificity. In February 2026, GRAIL and the NHS announced primary results from the 142,000-participant NHS-Galleri trial: a clinically meaningful reduction in stage IV diagnoses across those twelve cancers, with a greater-than-20% drop in the second and third annual rounds.

The patent record tracks this shift. US 12,410,480, granted September 2025, claims methods for colorectal cancer detection from cell-free nucleic acids using per-read methylation signals in identified genomic regions as input features to a machine-learning classifier. US 10,731,224 (granted August 2020) and its 2025 continuation US 12,234,515 cover cancer-screening enhancement via cell-free viral nucleic acids — turning reactivated viral fragments in blood into cancer signal. These are very different inventions than the Petricoin approach: sequencing rather than mass spectrometry, methylation and fragmentomics rather than SELDI peaks, and large validation cohorts rather than a single nonrandomized run.

Verdict: Wrong mechanism. The thesis — cancer detectable from blood-borne molecular patterns — is now demonstrably true at a scale Kurzweil’s 2005 reader would struggle to believe. The specific mechanism that was supposed to prove it in 2002 is gone.

Prediction 2: The Pap-smear pre-screener that grew into a thousand AI devices

Kurzweil’s second example is retrospectively true and, by today’s standards, radically understated. Neopath’s AutoPap 300QC was FDA-cleared in 1998. TriPath Imaging, formed by Neopath’s merger with AutoCyte in 1999, marketed it as FocalPoint. BD acquired TriPath in 2006. The BD FocalPoint GS Imaging System received FDA Premarket Approval in December 2008 after a trial on 12,732 SurePath slides across four US sites, and is still in routine use in cervical-cytology labs.

Kurzweil’s sentence is “such commercial products existed by 2003.” They did. What followed is where the scorecard gets interesting.

By late 2025, the FDA’s public list of AI/ML-enabled medical devices had crossed 1,356 cumulative authorizations, with 1,039 in radiology alone and imaging accounting for roughly 77% of all authorizations. The 2024 calendar year alone added 168 new authorizations — more than the total stock of such devices in 2018. The trajectory ran far past Kurzweil’s retrospective mention of a single cervical-cytology product.

The patent record mirrors the FDA curve. US 12,315,152 (granted May 2025) claims a three-dimensional deep-learning model trained on CT images paired with surgical-pathology outcome values to produce automated diagnoses of disease-database entities. US 12,051,200 (July 2024) covers an AI-based system that trains on annotated microscope images to issue diagnostic inferences on new images. US 11,568,538 (January 2023) covers a tumor-detection pipeline with preprocessing, segmentation, and classification stages. Each of these reads like the industrial generalization of FocalPoint’s original idea: take a visual diagnostic task, train a statistical model on labeled examples, and deploy it as a pre-read for a human specialist.

The clinical-trial record completes the picture. In the ClinicalTrials.gov registry, registered studies whose titles mention artificial intelligence, deep learning, machine learning, or neural network grew from 49 starts in 2018 to 279 in 2024, with another 269 already registered to start in 2025. The field of diagnostic AI did not stagnate after TriPath. It productized.

Verdict: Ahead of schedule. Kurzweil’s 2003 example was correct but conservative; the category it represented is now one of the best-documented commercialization curves in regulated medicine.

The scorecard

What Kurzweil got right, and what he missed

The pattern in this batch is a specific kind of forecasting error — the right bet on the wrong horse. Kurzweil in 2005 picked two exemplars to demonstrate that AI-driven diagnostics were not speculative fiction. Exemplar one turned out to be a statistical artifact that bankrupted the company trying to sell it. Exemplar two turned out to be the thin edge of a thousand-device wedge. He could not have known which would go which way from inside 2005. What he got right, in both cases, was the direction of the underlying tide. Blood-based molecular-pattern cancer detection happened; it just required the genomics revolution of the 2010s rather than the proteomics optimism of the early 2000s. AI-assisted imaging diagnostics happened; it just required convolutional networks rather than the hand-engineered feature extractors FocalPoint used.

This is the practical version of Kurzweil’s “direction is easier than timing” argument, and it cuts both ways. When he cites a specific company or a specific paper as proof-of-concept, the specific object is usually the most fragile part of the claim. The category survives; the exemplar often doesn’t. Readers looking at the 2024 book’s confidence that AI will reach “superhuman performance levels at virtually all diagnostic tasks” by the end of the 2020s should remember that the 2005 book’s two examples of 2002–2003 diagnostic AI had a 50% mortality rate at the exemplar level — and 100% survival at the thesis level.

Method note

We compared each prediction against the scientific literature, the US patent record, and the ClinicalTrials.gov registry using full-text indexes, then read the most important recent inventions and papers to understand their mechanisms. Patent numbers quoted above link to the actual grants. Citation counts are from the OpenAlex index of 357M scholarly works. Trial counts are from the full ClinicalTrials.gov registry. Web searches were used to confirm the historical record of the Petricoin controversy, the TriPath/BD acquisition timeline, the NHS-Galleri and PATHFINDER 2 results, and the current FDA AI/ML device count. Every number in this post came from a query or web fetch run during this session.