Search

We introduce LLMD, a large language model designed to analyze a patient's medical history based on their medical records. Along with domain knowledge, LLMD is trained on a large corpus of records collected over time and across facilities, as well as tasks and labels that make nuanced connections among them. This approach is critical to an accurate picture of patient health, and has distinctive advantages over models trained on knowledge alone, unlabeled records, structured EHR data, or records from a single health system. The recipe for LLMD continues pretraining a foundational model on both domain knowledge and the contents of millions of records. These span an average of 10 years of care and as many as 140 care sites per patient. LLMD is then instruction fine-tuned on structuring and abstraction tasks. The former jointly identify and normalize document metadata, provenance information, clinical named-entities, and ontology mappings, while the latter roll these into higher-level representations, such a continuous era of time a patient was on a medication. LLMD is deployed within a layered validation system that includes continual random audits and review by experts, e.g. based on uncertainty, disease-specific rules, or use-case. LLMD exhibits large gains over both more-powerful generalized models and domain-specific models. On medical knowledge benchmarks, LLMD-8B achieves state of the art accuracy on PubMedQA text responses, besting orders-of-magnitude larger models. On production tasks, we show that LLMD significantly outperforms all other models evaluated, and among alternatives, large general purpose LLMs like GPT-4o are more accurate than models emphasizing medical knowledge. We find strong evidence that accuracy on today's medical benchmarks is not the most significant factor when analyzing real-world patient data, an insight with implications for future medical LLMs.'

Data-driven algorithms play a large role in decision making across a variety of industries. Increasingly, these algorithms are being used to make decisions that have significant ramifications for people's social and economic well-being, e.g. in sentencing, loan approval, and policing. Amid the proliferation of such systems there is a growing concern about their potential discriminatory impact. In particular, machine learning systems which are trained on biased data have the potential to learn and perpetuate those biases. A central challenge for practitioners is thus to determine whether their models display discriminatory bias. Here we present a case study in which we frame the issue of bias detection as a causal inference problem with observational data. We enumerate two main causes of bias, sampling bias and label bias, and we investigate the abilities of six different fairness metrics to detect each bias type. Based on these investigations, we propose a set of best practice guidelines to select the fairness metric that is most likely to detect bias if it is present. Additionally, we aim to identify the conditions in which certain fairness metrics may fail to detect bias and instead give practitioners a false belief that their biased model is making fair decisions.

Engineered substrates offer a promising avenue towards graphene devices having tunable properties. In particular, topographically patterned substrates can expose unique behavior due to their ability to induce local variations in strain and electrostatic doping. However, to explore the range of possible science and applications, it is important to create topographic substrates which both have tunable features and are suitable for transport measurements. In this Letter we describe the fabrication of tunable, topographically patterned substrates suitable for transport measurements. We report both optical and transport measurements of graphene devices fabricated on these substrates, and demonstrate characteristic strain and local doping behavior induced by the topographic features.

A two-neutron unbound excited state of $^{24}$O was populated through a (d,d') reaction at 83.4 MeV/nucleon. A state at $E = 715 \pm 110$ (stat) $\pm 45 $ (sys) keV with a width of $\Gamma < 2$ MeV was observed above the two-neutron separation energy placing it at 7.65 $\pm$ 0.2 MeV with respect to the ground state. Three-body correlations for the decay of $^{24}$O $\rightarrow$ $^{22}$O + $2n$ show clear evidence for a sequential decay through an intermediate state in $^{23}$O. Neither a di-neutron nor phase-space model for the three-body breakup were able to describe these correlations.

The neutron-unbound nucleus 13Be was populated with a nucleon-exchange reaction from a 71 MeV/u secondary 13B beam. The decay energy spectrum was reconstructed using invariant mass spectroscopy based on 12Be fragments in coincidence with neutrons. The data could be described with an s-wave resonance at E = 0.73(9) MeV with a width of Gamma = 1.98(34) MeV and a d-wave resonance at E = 2.56(13) MeV with a width of Gamma = 2.29(73) MeV. The observed spectral shape is consistent with previous one-proton removal reaction measurements from 14B., Comment: Published in Phys. Rev. C

Graphene's linear dispersion relation and the attendant implications for bipolar electronics applications have motivated a range of experimental efforts aimed at producing p-n junctions in graphene. Here we report electrical transport measurements of graphene p-n junctions formed via simple modifications to a PbZr$_{0.2}$Ti$_{0.8}$O$_3$ substrate, combined with a self-assembled layer of ambient environmental dopants. We show that the substrate configuration controls the local doping region, and that the p-n junction behavior can be controlled with a single gate. Finally, we show that the ferroelectric substrate induces a hysteresis in the environmental doping which can be utilized to activate and deactivate the doping, yielding an `on-demand' p-n junction in graphene controlled by a single, universal backgate.

We examine the mechanical properties of graphene devices stretched on flexible elastomer substrates. Using atomic force microscopy, transport measurements, and mechanics simulations, we show that micro-rips form in the graphene during the initial application of tensile strain; however subsequent applications of the same tensile strain elastically open and close the existing rips. Correspondingly, while the initial tensile strain degrades the devices' transport properties, subsequent strain-relaxation cycles affect transport only moderately, and in a largely reversible fashion, yielding robust electrical transport even after partial mechanical failure.

A search for the neutron-unbound nucleus $^{21}$C was performed via the single proton removal reaction from a beam of 22 N at 68 MeV/u. Neutrons were detected with the Modular Neutron Array (MoNA) in coincidence with $^{20}$C fragments. No evidence for a low-lying state was found, and the reconstructed $^{20}$C+n decay energy spectrum could be described with an s-wave line shape with a scattering length limit of |as| < 2.8 fm, consistent with shell model predictions. A comparison with a renormalized zero-range three-body model suggests that $^{22}$C is bound by less than 70 keV., Comment: Accepted for publication in Nuclear Physics A

The ground state of neutron-rich unbound $^{13}$Li was observed for the first time in a one-proton removal reaction from $^{14}$Be at a beam energy of 53.6 MeV/u. The $^{13}$Li ground state was reconstructed from $^{11}$Li and two neutrons giving a resonance energy of 120$^{+60}_{-80}$ keV. All events involving single and double neutron interactions in the Modular Neutron Array (MoNA) were analyzed, simulated, and fitted self-consistently. The three-body ($^{11}$Li+$n+n$) correlations within Jacobi coordinates showed strong dineutron characteristics. The decay energy spectrum of the intermediate $^{12}$Li system ($^{11}$Li+$n$) was described with an s-wave scattering length of greater than -4 fm, which is a smaller absolute value than reported in a previous measurement., Comment: Accepted for publication in Phys. Rev. C as a Rapid Communication

The technique of invariant mass spectroscopy has been used to measure, for the first time, the ground state energy of neutron-unbound $^{28}\textrm{F},$ determined to be a resonance in the $^{27}\textrm{F} + n$ continuum at $2\underline{2}0 (\underline{5}0)$ keV. States in $^{28}\textrm{F}$ were populated by the reactions of a 62 MeV/u $^{29}\textrm{Ne}$ beam impinging on a 288 $\textrm{mg/cm}^2$ beryllium target. The measured $^{28}\textrm{F}$ ground state energy is in good agreement with USDA/USDB shell model predictions, indicating that $pf$ shell intruder configurations play only a small role in the ground state structure of $^{28}\textrm{F}$ and establishing a low-$Z$ boundary of the island of inversion for N=19 isotones., Comment: 5 pages, 4 figures, to be published in Phys. Rev. Lett

Data from the operation of a bubble chamber filled with 3.5 kg of CF$_{3}$I in a shallow underground site are reported. An analysis of ultrasound signals accompanying bubble nucleations confirms that alpha decays generate a significantly louder acoustic emission than single nuclear recoils, leading to an efficient background discrimination. Three dark matter candidate events were observed during an effective exposure of 28.1 kg-day, consistent with a neutron background. This observation provides the strongest direct detection constraint to date on WIMP-proton spin-dependent scattering for WIMP masses $>20$ GeV/c$^{2}$., Comment: 4 pages, 4 figures V2 submitted to match journal version

The first excited state in neutron-rich 23O was observed in a (2p1n) knock-out reaction from 26Ne on a beryllium target at a beam energy of 86 MeV/A. The state is unbound with respect to neutron emission and was reconstructed from the invariant mass from the 22O fragment and the neutron. It is unbound by 45(2) keV corresponding to an excitation energy of 2.8(1) MeV. The non-observation of further resonances implies a predominantly direct reaction mechanism of the employed three-nucleon-removal reaction which suggests the assignment of the observed resonance to be the 5/2+ hole state., Comment: 7 pages, 3 figures, Proc. 23rd Winter Workshop on Nuclear Dynamics, February 11-18, 2007

The two-proton knockout reaction 9Be(26Ne,O2p) was used to explore excited unbound states of 23O and 24O. In 23O a state at an excitation energy of 2.79(13) MeV was observed. There was no conclusive evidence for the population of excited states in 24O., Comment: 6 pages, 3 figures, Proc. 9th Int. Spring Seminar on Nucl. Phys. Changing Facets of Nuclear Structure, May 20-34, 2007

The Modular Neutron Array (MoNA) was used in conjunction with a large-gap dipole magnet (Sweeper) to measure neutron-unbound states in oxygen isotopes close to the neutron dripline. While no excited states were observed in 24O, a resonance at 45(2) keV above the neutron separation energy was observed in 23O., Comment: 6 pages, 4 Figures, submitted to Proc. Int. Conf. on Proton Emitting Nuclei and Related Topics, PROCON07

We have observed an excited state in the neutron-rich semi-magic nucleus O-23 for the first time. No such states have been found in previous searches using gamma-ray spectroscopy. The observation of a resonance in n-fragment coincidence measurements confirms the speculation in the literature that the lowest excited state is neutron unbound and establishes positive evidence for a 2.8(1) MeV excitation energy of the first excited state in O-23. The non-observation of a predicted second excited state is explained assuming selectivity of inner-shell knockout reactions., Comment: 4 pages including 4 figures