Vetting an opposing medical expert in 4 numbers

The four numbers

Running case for the rest of this post: a New York wrong-site / total-knee-arthroplasty matter, plaintiff side, sourcing an orthopedic expert, then pre-vetting the orthopaedic surgeon the defense just disclosed as their own expert.

1. Willingness Score

The plaintiff problem: you need an orthopedic surgeon who will actually take a wrong-site total-knee-arthroplasty case in New York. Most high-volume joint surgeons will not. They operate alongside the people they would have to testify against, and they know it. So you start dialing, and you can burn a week of afternoons collecting polite nos. The Willingness Score exists to tell you who is likely to say yes before you pick up the phone.

The score is a 0 to 100 composite with a letter grade, built from nine signals:

• Prior testifying appearances in the corpus (0 to 50 pts)
• Recency of the last appearance (−5 to +25)
• Daubert exposure, i.e. has been challenged before (+5)
• NPI active or deactivated (−25 to +5)
• Specialty match to the allegation area (−10 to +15)
• On a state IME / QME roster (+10)
• Recent PubMed publications in the area (+5 to +15)
• Compliance flags (−25 to +5)
• Industry consulting receipts on Open Payments (0 to +10)

The candidate pool is built two ways. PubMed authors of articles whose title or abstract matches the allegation phrase over the last ten years. Plus named experts in the corpus whose case captions mention the allegation. Then every candidate is resolved to an NPI, and any candidate whose primary NUCC taxonomy code does not match the inferred allegation specialty is dropped from the pool. A physical therapist or a case manager never surfaces as an orthopedic-surgery candidate.

The phrasing matters. Query “wrong-site knee replacement” and the pool resolves to patient-safety and review-literature authors, not treating surgeons — all 60 candidates get gated out and the system hands back a recovery hint telling you to re-run with the clinical procedure term. Query “total knee arthroplasty” instead and the top of the list looks like this:

What the result actually tells you. Candidate A is the top score, but practices in Minnesota. The New York match, Candidate D, sits at #4 with a High grade, which is the venue the plaintiff actually needs. The state field is not used to filter the candidate pool, because the best expert for a venue may practice elsewhere; the score is venue-agnostic and you read it as an information list, not as a pick.

Each candidate ships with a score breakdown showing exactly which signals contributed how many points. For Candidate A the breakdown reads: prior_appearances +50 from 12 prior case mentions in the corpus, specialty_fit +15 for an Orthopaedic Surgery taxonomy match, npi_active +5, publications_in_area +10 for 108 matching articles, and recency pulled to 0 because his last appearance on record was 2017. No IME bonus because we did not match him to a state IME roster; no industry-consulting hit on this candidate. The number is built, you can audit it line by line, and you can argue with it.

2. Vulnerability Score

Once a candidate is named, the next question is Daubert risk. The Vulnerability Score is the descriptive ratio of excluded testimony to total Daubert challenges across both expert corpora:

• The medical-expert corpus, 206K federal and appellate cases, where a daubert_challenge tag flags rows that carry no admit / exclude outcome on their own; they are the “has been challenged” flag plus the matched case list.
• The damages-expert corpus, where is_daubert and daubert_outcome (excluded / admitted / challenged) are extracted from the opinion text. This is where the actual vulnerability percentage comes from.

Run on the disclosed defense expert, the system returns: 0 prior case mentions in the 206K-case medical corpus, 0 in the state-trial corpus, 0 Daubert events of any kind — against 38 PubMed articles on knee arthroplasty and ACL reconstruction. Vulnerability comes back null because the denominator is zero, and we would rather show you that than invent a number. He is a heavily-published surgeon who has never been Daubert-challenged on record, so there is no exclusion rate to compute.

Where the score earns its keep: surveying a list of possible damages-side experts that opposing counsel has used in the same venue before. The same forensic economist who appears in three of the last five med-mal cases against your hospital is exactly the person you want to see the Vulnerability column on.

3. Side Lean

The deposition question every partner asks: which side does this expert testify for, historically? VerdictSearch and JuryVerdictReporter answer this by inferring from verdict outcome, which is the weakest possible proxy. The Side Lean classifier answers it by reading the opinion text itself.

For a single opinion where the expert is named, the answer comes back as a plain-language read with the matched evidence underneath:

Run across every published opinion the expert appears in, the aggregate looks like “defense in 7 of 9 prior appearances, with two court-appointed neutral roles.” That is a real fact, sourced from the opinions themselves, not from a verdict-report inference. For an expert who advertises as “available to plaintiffs and defense alike,” this is the answer that ends the marketing version of the story.

On the defense expert, the classifier returns nothing — he has zero opinions in either corpus, so there is no language to read and the side-lean object comes back empty. That is not a defect; it is the same finding as the Vulnerability null, read from the other direction. A surgeon with 38 publications and no courtroom paper trail does not have a provable side lean, and you plan the deposition around that fact rather than around an inferred one.

Caveat: the classifier reads opinion text with a regex-pattern matcher, not a fully trained model. Confidence scores below 0.5 should be read as “the opinion did not give a clean signal,” not as evidence of court-appointed neutrality. The snippet field always ships so you can read the matched language yourself.

4. Shared Sponsor overlap

Here is where the running case turns. Case count, Daubert exposure, and side lean all came back empty on this expert — a heavily-published academic with no courtroom record hands you nothing to impeach on the usual axes. The money is the exception, and it is where this expert is wide open.

The fourth number is the one no general-purpose legal platform ships. Pass the opposing expert and the defendant doctor; the system returns every device or pharma manufacturer that paid both, on Open Payments, with dollars, years, and product names. Subsidiaries collapse to corporate parent through a curated 50-family alias table:

• Davol, Bard Peripheral Vascular, BD → BD
• Covidien, Medtronic Vascular → Medtronic
• Ethicon, DePuy Synthes, Medical Device Business Services → J&J MedTech
• Allergan, AbbVie Aesthetics, Allergan Aesthetics → AbbVie

Run on the matchup we have been working through, this expert is the opposite of a clean one. His industry footprint is $6,981,067 in general payments across 371 records, and $6,417,322 of it — roughly 92% — comes from a single manufacturer, Smith & Nephew, concentrated in a $5,792,687 spike in 2022 alone. That is the kind of number a jury hears.

Cross-referenced against the defendant surgeon, six corporate parents paid both of them. The cross-exam question writes itself. “Doctor, you and the defendant both took money from Exactech for the OPTETRAK knee system in 2017, 2018, and 2019 — $161,199 to you, $35,350 to him. Were you at the same industry-sponsored events?” The answer is almost always no, and the foundation for impeachment is built into the exhibit.

Honest limits

Three things you should know before you build a workflow on top of these scores.

1. 01Names are not NPIs. When you query by name on a common surname, the system tries to disambiguate via NPI registry + first initial, but a “John Chan” query in California can resolve to a different John Chan than the one you want. Pass NPI explicitly when you have it. The dossier also returns alternate candidates so you can pick the right one.
2. 02The corpus is appellate-skewed. An expert who has testified in 25 state-trial cases that never produced a published appellate opinion will appear as “light” in our case count. The Willingness Score handles this by weighing recency and publication footprint alongside corpus appearances, but if the only thing you have on an expert is corpus appearances, treat it as a lower bound.
3. 03Vulnerability is descriptive, not predictive. The score is excluded / total, not the probability the expert will be excluded in your case. Sample sizes under five Daubert events should be read as directional only. The per-case list ships with case_name, court, decision_date, and opinion_url so you read the actual rulings.

How to use it

The four scores ship behind one chat conversation. A typical plaintiff workflow opens with the Willingness Score on the allegation area to source a candidate, then the Vulnerability Score and Side Lean on the named opposing expert once the answer arrives. A typical defense workflow runs the Shared Sponsor overlay between the named defendant and the plaintiff's expert the moment the expert disclosure lands.

All four are also callable from Claude, ChatGPT, or any agent that speaks MCP. The agent surface is the same as the chat surface, so a firm with an internal AI workflow already wired can hit the scores from their own assistant. From $199 a month, billed to client matter.

For the broader product surface around this, the med-mal page covers the NPDB cohort percentile bands, the fact-pattern to standard-of-care pairing, the 51-state damage cap and statute of limitations lookup, and the 16 carrier closed-claims studies we've indexed.