Snigdha Banda

0 followers · 265 views

on the atlas · 30

highlights · 137

What I know now is that the greatest killer of momentum isn’t the wrong path, it’s not getting started.
making the decision right - by Sarah - liminal space
Since I moved to Spain, I have made every decision with faith and intuition and some have worked better than others.
making the decision right - by Sarah - liminal space
For example, Is- lam and Hinduism do not consider aggression and femi- ninity as antithetical
AR_SV Indian Country
In most cases women are able to neither control vio- lence against themselves nor modify their abusers’ behaviors according to their own will
AR_SV Indian Country
Men, on the other hand, often minimize their violence against female partners or blame the victim, which reflects a greater sense of entitlement to such behavior than for women.
AR_SV Indian Country
which is based on establishing mastery, su- premacy, and authority
AR_SV Indian Country
demanding attention, expressing anger, escape, and punishment as motives that compel women to engage in violent behavior.
AR_SV Indian Country
On the CTS, both behaviors would be assessed at a lower magnitude of violence than physical hits.
AR_SV Indian Country
The dynamics of domestic violence involve the goal of dominating women by uti- lizing various tactics of coercive control in both public and private arenas so as to maintain the systems of patriarchy in society.
AR_SV Indian Country
The main critique of studies using the CTS cen- ters on the argument that the scales do not allow any room for contexts and motives of intimate partner vio- lence
AR_SV Indian Country
Some research- ers and activists tend to define battering as a pattern of intimidation, coercive control, and oppression
AR_SV Indian Country
To her, one’s heritage and the effects of colonialism were universal. It governed relations on both interpersonal and macro levels.
Shamita Das Dasgupta – Badass Asian Americans
Cerner and Epic each currently charge $5,000 for a single app registration and another $5,000 for each additional app
Developers complain of high EHR fees for SMART apps - POLITICO
corticotropin-releasing factor (CRF)/hypothalamic-pituitary-adrenal (HPA) axis system plays an important role in the stress response. CRF is released from the hypothalamus, with stimulation of adrenocorticotropic hormone (ACTH) release from the pituitary, resulting in glucocorticoid (Cortisol in man) release from the adrenal, which in turn has a negative feedback effect on the axis at the level of the pituitary, as well as central brain sites including hypothalamus and hippocampus.
Traumatic stress: effects on the brain - PMC
You want to alternate between being a mirror and an excavator, to help them understand themselves better and come to a solution on their own.
How to save your friends - by Kasra - Bits of Wonder
The more obvious the solution seems to you, the more you should resist prescribing it, unless you’re sure that they actually haven’t considered it before.
How to save your friends - by Kasra - Bits of Wonder
If your goal is to help them, try to maintain the center of gravity on them.
How to save your friends - by Kasra - Bits of Wonder
You should view such conversations as an excavation of your friend’s mind.
How to save your friends - by Kasra - Bits of Wonder
Drugs that block the stimulating effects of acetylcholine (the anticholinergics) as well as chemicals that mimic the effects of acetylcholine, such as the organophosphate insecticides, can impair the ability to think
New Page Title Here
The parasympathetic, cholinergic, nervous system had been thought of as inactive during stress, and activated to regulate process
New Page Title Here
it’s sharing these things as they are happening, expressing your anger and anxiety and sadness while you still haven’t resolved them.
Give your friends a chance to abandon you - by Kasra
s BigQuery and Looker
care.ai Builds Advanced Solutions for Nurses with Google Cloud's Generative AI and Data Analytics on Its Smart Care Facility Platform - Oct 9, 2023
help reduce administrative burdens, mitigate staffing shortages, and free up clinicians to spend more time with patients in 1,500 acute and post-acute facilities where care.ai's platform is already deployed
care.ai Builds Advanced Solutions for Nurses with Google Cloud's Generative AI and Data Analytics on Its Smart Care Facility Platform - Oct 9, 2023
he was a bit like a chatty lawyer, and a bit like a surrealist comic
Life update, March - by Sasha Chapin - Sasha's 'Newsletter'
each additional 15 minutes of total daily EHR time was associated with 0.58 (95% CI, 0.32-0.84) percentage point greater panel-level hemoglobin A1c control, 0.52 (95% CI, 0.33-0.71) percentage point greater hypertension control, and 0.28 (95% CI, 0.05-0.52) higher breast cancer screening rates
Association Between Electronic Health Record Time and Quality of Care Metrics in Primary Care - PubMed
significant associations between EHR time and panel-level achievement of hemoglobin A1c control, hypertension control, and breast cancer screening targets
Association Between Electronic Health Record Time and Quality of Care Metrics in Primary Care - PubMed
daily total EHR time, after-hours time, time from 5:30 pm to 7:00 am and time on weekends, and daily EHR time on notes, sending and receiving patient, staff, results, prescription, or system messages [in-basket], and clinical review
Association Between Electronic Health Record Time and Quality of Care Metrics in Primary Care - PubMed
CMS estimates that roughly 485,000 physicians in total could ultimately be eligible to participate.
The Federal Government Has Put Billions into Promoting Electronic Health Record Use: How Is It Going? | Commonwealth Fund
At a minimum, that will mean having systems capable of e-prescribing, reporting quality data, and exchanging data among providers.
The Federal Government Has Put Billions into Promoting Electronic Health Record Use: How Is It Going? | Commonwealth Fund
Weak Supervision For weak supervision, we only consider the 97% of the dataset where the overlap with an answer choice was at least 5 char- acters as candidates for pseudolabels. Following prior work (Lang et al., 2022a,b), we additionally used a technique called the cut statistic to select a high-quality subset of the weakly labeled data to re- duce the noise in the training process. We selected a subset of size 75% to decrease noise while still choosing a large enough set to ensure all acronyms were seen during training. We fine-tuned a Pub- MedBERT (Gu et al., 2021) model, a BERT variant th…
Large language models are few-shot clinical information extractors
fter annotation of the data, we create three versions of the dataset: token-level, phrase-level, and relation-level. For the first, we split all word in the example and assigned them their respective label or none if they were not part of a label (
Large language models are few-shot clinical information extractors
Guidelines excluded medication categories (e.g. “ACE-inhibitor”) if they referred to more specific drug names mentioned elsewhere
Large language models are few-shot clinical information extractors
Accuracy and macro F1 for zero-shot language modeling
Large language models are few-shot clinical information extractors
nd show that we can instead view the LLM + resolver system as a labeler rather than as a classifier , and that this can even boost performance.
Large language models are few-shot clinical information extractors
Since we cannot query GPT-3 on this dataset, we distill and transfer a model trained on the outputs from Dataset 1.
Large language models are few-shot clinical information extractors
On the tasks below, we find that GPT-3 + R matches or exceeds strong few- shot, zero-shot, and even supervised baselines.
Large language models are few-shot clinical information extractors
While CASI was originally annotated for acronym disambiguation, we created three new an- notated datasets from existing snippets of the CASI dataset
Large language models are few-shot clinical information extractors
s seen in Figure 1, this consists of (i) a one-shot exam- ple with an output in the desired structured format (which could be incorrect content-wise), and (ii) guiding the model to use the same format. Specific constructions are found in Sections 6 and 7.
Large language models are few-shot clinical information extractors
¹f 𝑥 𝑖 𝑎 𝑖 gº 𝑛 𝑖 = 1 , where 𝑥 𝑖 is the input text as a string, 𝑎 𝑖 is (optional) side information as a string (e.g., which acronym to disambiguate).
Large language models are few-shot clinical information extractors
However, in several applications, researchers observed the performance gains to be marginal to none over clas- sical methods such as logistic regression
Large language models are few-shot clinical information extractors
While InstructGPT (Ouyang et al., 2022) has 2% or 1 𝑘 extraction examples in its training, the LLM output is never converted to a structured form, and extraction examples are only evaluated qualitatively for improvement over other models. That is, only results for classification and genera- tion tasks are quantified.
Large language models are few-shot clinical information extractors
More recently, large language models such as T0 and InstructGPT have re-configured their training objectives to explic- itly encourage the model to perform well at such prompts
Large language models are few-shot clinical information extractors
We therefore introduce guided prompt design to steer the LLM towards an easy- to-structure output and resolvers to map from the LLM outputs to the structured label space; see Figure 1.
Large language models are few-shot clinical information extractors
we fine-tune PubMedBERT (Gu et al., 2021) and follow Lang et al. (2022a); details and hyperparameters are found in the appendix.
Large language models are few-shot clinical information extractors
We used GPT-3 edit (using engine text-davinci-edit-001 ) with greedy decoding (temperature = 0 ). For each exam- ple, we provided the full clinical snippet and appended the single instruction Expand the abbreviation:{abbr} . Since we did not provide the LLM with the answer choices, the form of the output string could still differ slightly from all the candidate answer
Large language models are few-shot clinical information extractors
erated from unlabeled text by replacing expansions (e.g. physical therapy ) with their acronyms ( PT ) and using the original expansion as the label.
Large language models are few-shot clinical information extractors
resolvers for guided prompts are much eas- ier to write than resolvers for un-guided prompts.
Large language models are few-shot clinical information extractors
de-identified Clinical Acronym Sense Inventory (CASI) dataset
Large language models are few-shot clinical information extractors
we handcraft our prompt templates using a set of 5 validation examples per task
Large language models are few-shot clinical information extractors
More recently, fol- lowing BERT, many clinical and biomedical vari- ations swiftly followed including ClinicalBERT, SciBERT, BioBERT, and PubMedBERT
Large language models are few-shot clinical information extractors