Try a new search

Format these results:

Searched for:



Total Results:


Health system-scale language models are all-purpose prediction engines

Jiang, Lavender Yao; Liu, Xujin Chris; Nejatian, Nima Pour; Nasir-Moin, Mustafa; Wang, Duo; Abidin, Anas; Eaton, Kevin; Riina, Howard Antony; Laufer, Ilya; Punjabi, Paawan; Miceli, Madeline; Kim, Nora C; Orillac, Cordelia; Schnurman, Zane; Livia, Christopher; Weiss, Hannah; Kurland, David; Neifert, Sean; Dastagirzada, Yosef; Kondziolka, Douglas; Cheung, Alexander T M; Yang, Grace; Cao, Ming; Flores, Mona; Costa, Anthony B; Aphinyanaphongs, Yindalon; Cho, Kyunghyun; Oermann, Eric Karl
Physicians make critical time-constrained decisions every day. Clinical predictive models can help physicians and administrators make decisions by forecasting clinical and operational events. Existing structured data-based clinical predictive models have limited use in everyday practice owing to complexity in data processing, as well as model development and deployment1-3. Here we show that unstructured clinical notes from the electronic health record can enable the training of clinical language models, which can be used as all-purpose clinical predictive engines with low-resistance development and deployment. Our approach leverages recent advances in natural language processing4,5 to train a large language model for medical language (NYUTron) and subsequently fine-tune it across a wide range of clinical and operational predictive tasks. We evaluated our approach within our health system for five such tasks: 30-day all-cause readmission prediction, in-hospital mortality prediction, comorbidity index prediction, length of stay prediction, and insurance denial prediction. We show that NYUTron has an area under the curve (AUC) of 78.7-94.9%, with an improvement of 5.36-14.7% in the AUC compared with traditional models. We additionally demonstrate the benefits of pretraining with clinical text, the potential for increasing generalizability to different sites through fine-tuning and the full deployment of our system in a prospective, single-arm trial. These results show the potential for using clinical language models in medicine to read alongside physicians and provide guidance at the point of care.
PMID: 37286606
ISSN: 1476-4687
CID: 5536672

Quantitative and Qualitative Evaluation of Provider Use of a Novel Machine Learning Model for Favorable Outcome Prediction

Yang, Elisabeth; Aphinyanaphongs, Yin; Punjabi, Paawan V; Austrian, Jonathan; Wiesenfeld, Batia
Predictive models may be particularly beneficial to clinicians when they face uncertainty and seek to develop a mental model of disease progression, but we know little about the post-implementation effects of predictive models on clinicians' experience of their work. Combining survey and interview methods, we found that providers using a predictive algorithm reported being significantly less uncertain and better able to anticipate, plan and prepare for patient discharge than non-users. The tool helped hospitalists form and develop confidence in their mental models of a novel disease (Covid-19). Yet providers' attention to the predictive tool declined as their confidence in their own mental models grew. Predictive algorithms that not only offer data but also provide feedback on decisions, thus supporting providers' motivation for continuous learning, hold promise for more sustained provider attention and cognition augmentation.
PMID: 37128409
ISSN: 1942-597x
CID: 5542392

Could the cure be the cause? Cefepime induced encephalopathy in a hospitalized older adult [Meeting Abstract]

Zweig, Y; Punjabi, P
Background: Delirium occurs commonly in hospitalized older adults. Clinicians investigate frequent causes of delirium such as infection, metabolic derangement, intracranial disorders, urinary and fecal disorders, and medication. The complexity in treating the underlying cause of delirium is the treatment of one contributing factor (e.g. infection), may in fact be the source of the confusion. Case report: We evaluated a 95 year old woman with atrial fibrillation, severe aortic stenosis, chronic kidney disease, heart failure, and clostridium difficile who presented to the hospital with shortness of breath. White blood cell count was elevated and CT chest revealed moderate patchy consolidation in the lung. The patient was started on Vancomycin, Aztreonam, and Cefepime for multidrug resistant pneumonia. At baseline she has no dementia and there was no concern for delirium on presentation. Hospital days 4-6 she became lethargic but remained clinically stable without a change in infectious markers. By hospital day 8 she completed the course of Cefepime but was nonverbal and opened her eyes only when directly stimulated. The only notable lab change was an uptrend of creatinine from 1.09 to 1.70 in the setting of diuresis for fluid overload. Mental status remained poor through hospital day 10. CT brain revealed moderate global volume loss. There was no evidence of a new infection, hypercarbia, or other metabolic derangements. EEG was not completed per family preference. By hospital day 12, 4 days after the cessation of Cefepime, her mental status began to improve and returned almost back to baseline before discharge.
Discussion(s): Cefepime is known to have neurotoxic effects that may include depressed consciousness, encephalopathy, aphasia, myoclonus, seizures, and coma. The primary risk factors for Cefepime neurotoxicity are renal impairment and blood-brain barrier dysfunction from systemic inflammation. Older age is a commonly reported clinical risk factor. Neurotoxic effects are noted a median of 4 days after initiation with reduced consciousness most commonly seen. The neurotoxic symptoms may resolve a median of 2 days after discontinuation. The literature findings mimic what was seen in this case. This case reinforces the need to consider the multifactorial contribution to delirium etiology. Cefepime induced neurotoxicity should be considered for infected patients with persistent hypoactive delirium
ISSN: 1532-5415
CID: 4757582

Does Scheduling a Postdischarge Visit with a Primary Care Physician Increase Rates of Follow-up and Decrease Readmissions?

Marcondes, Felippe O; Punjabi, Paawan; Doctoroff, Lauren; Tess, Anjala; O'Neill, Sarah; Layton, Timothy; Quist, Kramer; Mehrotra, Ateev
BACKGROUND:Driven in part by Medicare's Hospital Readmissions Reduction Program, hospitals are focusing on improving the transition from inpatient to outpatient care with particular emphasis on early follow-up with a primary care physician (PCP). OBJECTIVE:To assess whether the implementation of a scheduling assistance program changes rates of PCP follow-up or readmissions. DESIGN/METHODS:Retrospective cohort study. SETTING/METHODS:An urban tertiary care center PATIENTS: A total of 20,918 adult patients hospitalized and discharged home between September 2008 and October 2015. INTERVENTION/METHODS:A postdischarge appointment service to facilitate early PCP follow-up. MAIN MEASURES/METHODS:Primary outcomes were rates of follow-up visits with a PCP within seven days of discharge and hospital readmission within 30 days of discharge. Our first analysis assessed differences in outcomes among patients with and without the use of the service. In a second analysis, we exploited the fact that the service was not available on weekends and conducted an instrumental variable analysis that used the interaction between the intervention and day of the week of admission. RESULTS:In our multivariable analysis, use of the appointment service was associated with much higher rates of PCP follow-up (+31.9 percentage points, 95% CI: 30.2, 33.6; P < .01) and a decrease in readmission (-3.8 percentage points, 95% CI: -5.2, -2.4; P < .01). In the instrumental variable analysis, use of the service also increased the likelihood of a PCP follow-up visit (33.4 percentage points, 95% CI: 7.9, 58.9; P = .01) but had no significant impact on readmissions (-2.5 percentage points, 95% CI: -22.0, 17.0; P = .80). CONCLUSIONS:The postdischarge appointment service resulted in a substantial increase in timely PCP followup, but its impact on the readmission rate was less clear.
PMID: 31532749
ISSN: 1553-5606
CID: 4098032

Review of gastroesophageal reflux disease (GERD) in the diabetic patient

Punjabi, Paawan; Hira, Angela; Prasad, Shanti; Wang, Xiangbing; Chokhavatia, Sita
This article reviews the known pathophysiological mechanisms of comorbid gastroesophageal reflux disease (GERD) in the diabetic patient, discusses therapeutic options in care, and provides an approach to its evaluation and management. We searched for review articles published in the past 10 years through a PubMed search using the filters diabetes mellitus, GERD, pathophysiology, and management. The search only yielded a handful of articles, so we independently included relevant studies from these review articles along with related citations as suggested by PubMed. We found diabetic patients are more prone to developing GERD and may present with atypical manifestations. A number of mechanisms have been proposed to elucidate the connection between these two diseases. Studies involving treatment options for comorbid disease suggest conflicting drug-drug interactions. Currently, there are no published guidelines specifically for the evaluation and management of GERD in the diabetic patient. Although there are several proposed mechanisms for the higher prevalence of GERD in the diabetic patient, this complex interrelationship requires further research. Understanding the pathophysiology will help direct diagnostic evaluation. In our review, we propose a management algorithm for GERD in the diabetic patient.
PMID: 25706050
ISSN: 1753-0407
CID: 2544492

Memorising Milton's Paradise lost: a study of a septuagenarian exceptional memoriser [Case Report]

Seamon, John G; Punjabi, Paawan V; Busch, Emily A
At age 58, JB began memorising Milton's epic poem Paradise Lost. Nine years and thousands of study hours later, he completed this process in 2001 and recalled from memory all 12 books of this 10,565-line poem over a 3-day period. Now 74, JB continues to recite this work. We tested his memory accuracy by cueing his recall with two lines from the beginning or middle of each book and asking JB to recall the next 10 lines. JB is an exceptional memoriser of Milton, both in our laboratory tests in which he did not know the specific tests or procedures in advance, and in our analysis of a videotaped, prepared performance. Consistent with deliberate practice theory, JB achieved this remarkable ability by deeply analysing the poem's structure and meaning over lengthy repetitions. Our findings suggest that exceptional memorisers such as JB are made, not born, and that cognitive expertise can be demonstrated even in later adulthood.
PMID: 20419555
ISSN: 1464-0686
CID: 2544502

Does feigning amnesia impair subsequent recall?

Sun, Xue; Punjabi, Paawan V; Greenberg, Lucy T; Seamon, John G
Defendants who are accused of serious crimes sometimes feign amnesia to evade criminal responsibility. Previous research has suggested that feigning amnesia might impair subsequent recall. In two experiments, participants read and heard a story about a central character, described as "you," who was responsible for the death of either a puppy (Experiment 1) or a friend (Experiment 2). On free and cued recall tests immediately after the story, participants who had feigned amnesia recalled less than did participants who had recalled accurately. One week later, when all participants recalled accurately, participants who had previously feigned amnesia still performed worse than did participants who had recalled accurately both times. However, the participants who had formerly feigned amnesia did not perform worse than did a control group who had received only the delayed recall tests. Our results suggest that a "feigned amnesia effect" may reflect nothing more than differential practice at recall. Feigning amnesia for a crime need not impair memory for that crime when a person later seeks to remember accurately.
PMID: 19103978
ISSN: 0090-502x
CID: 2544512