Skip to content

Research

Every number we cite, and where it came from.

This site cites published research about how hospitals and their systems behave. Each figure carries a superscript, and each superscript lands here, beside the study it came from and a link to read it in full. Nothing on this page is a result at a hospital running Pensieve.

What belongs on this page

A figure may be cited anywhere on this site only if it appears here, with a public source a reader can open and check. That is the whole rule.

It applies to us as much as to anyone quoting us.

These studies describe the problem. They were conducted on hospitals and clinics that had never heard of Pensieve, and they are cited to show that the problems this platform is aimed at are real and measured. None of them says we have solved anything for anyone. What Pensieve does about them is stated as mechanism on the architecture pages.

Pensieve outcome measures on this page: none

Hospitals run Pensieve today, under NDA. No figure below is ours, and none of them is evidence about our software.

Numbered as they are cited. Each entry prints the sentence the site is allowed to use, then what the study measured underneath it.

Sources

The list

  1. Across four specialties, physicians spent 49.2 percent of the office day on the record and desk work, and 27.0 percent in direct time with patients.Direct observation of 57 physicians in family medicine, internal medicine, cardiology, and orthopaedics. Of the office day, 49.2 percent went to the electronic record and desk work against 27.0 percent of direct clinical face time with patients. In the examination room itself, 37.0 percent of the time still went to the record and desk work. Sinsky C, Colligan L, Li L, et al. Allocation of Physician Time in Ambulatory Practice: A Time and Motion Study in 4 Specialties. Annals of Internal Medicine, 2016;165(11):753-760.
    (opens in a new tab)
  2. In the trial that established computerised order entry, serious medication errors that reached the patient fell by 55 percent.A controlled trial in an academic tertiary hospital. Non-intercepted serious medication errors fell 55 percent, from 10.7 to 4.86 events per 1,000 patient-days, after physician order entry was introduced. Bates DW, Leape LL, Cullen DJ, et al. Effect of Computerized Physician Order Entry and a Team Intervention on Prevention of Serious Medication Errors. JAMA, 1998;280(15):1311-1316.
    (opens in a new tab)
  3. A National Academies review concluded that most people will experience at least one diagnostic error in their lifetime, and named information that fails to reach the decision as a recurring contributor.A consensus study of the diagnostic process. It concluded that most people will receive at least one wrong or delayed diagnosis in their lifetime, and identified health information technology that does not support the diagnostic process, together with poor information transfer across care settings, as contributing factors. National Academies of Sciences, Engineering, and Medicine. Improving Diagnosis in Health Care. Washington, DC: The National Academies Press, 2015.
    (opens in a new tab)
  4. Across 73 studies, direct communication between hospital and primary care physicians occurred in 3 to 20 percent of cases, and a discharge summary was available at the first follow-up visit 12 to 34 percent of the time.A systematic review of 73 studies of the hospital to primary care handover. Direct communication between the two physicians happened in 3 to 20 percent of cases. A discharge summary had reached the primary care physician by the first post-discharge visit in 12 to 34 percent of cases, and it was frequently missing the diagnostic test results and the pending results at discharge. Kripalani S, LeFevre F, Phillips CO, et al. Deficits in Communication and Information Transfer Between Hospital-Based and Primary Care Physicians. JAMA, 2007;297(8):831-841.
    (opens in a new tab)
  5. After nine hospitals put the handover between clinical teams into one structured form, medical errors fell 23 percent and preventable adverse events fell 30 percent across 10,740 admissions.A prospective study at nine paediatric residency programmes, comparing the six months before and the six months after a single structured verbal and written handover was adopted. In 10,740 admissions the medical error rate fell from 24.5 to 18.8 per 100 admissions and preventable adverse events from 4.7 to 3.3 per 100 admissions, both at P less than 0.001. The duration of the handover itself was unchanged. Starmer AJ, Spector ND, Srivastava R, et al. Changes in Medical Errors after Implementation of a Handoff Program. New England Journal of Medicine, 2014;371(19):1803-1812.
    (opens in a new tab)
  6. At eight hospitals across eight countries, deaths fell from 1.5 percent to 0.8 percent and inpatient complications from 11.0 percent to 7.0 percent after the surgical team began briefing off one checklist.A prospective study of 3,733 consecutive patients before and 3,955 after a nineteen item checklist was introduced at eight hospitals, one of them in New Delhi. The checklist obliges the surgeon, the anaesthetist and the nursing team to confirm identity, site, allergies, airway risk, antibiotic timing, imaging, blood loss and instrument counts aloud together at three points in the operation. The death rate was 1.5 percent at baseline and 0.8 percent afterwards, at P equal to 0.003; inpatient complications were 11.0 percent and 7.0 percent, at P less than 0.001. Haynes AB, Weiser TG, Berry WR, et al. A Surgical Safety Checklist to Reduce Morbidity and Mortality in a Global Population. New England Journal of Medicine, 2009;360(5):491-499.
    (opens in a new tab)
  7. Under a state mandate that put emergency sepsis care on one clock, each additional hour taken to complete the three-hour bundle carried higher risk-adjusted in-hospital mortality, at an odds ratio of 1.04 per hour across 49,331 patients at 149 hospitals.A retrospective study of 49,331 patients at 149 hospitals reporting to New York State under regulations that require blood cultures, a serum lactate and broad spectrum antibiotics within three hours of a sepsis protocol starting. The three-hour bundle was completed within three hours for 82.5 percent of patients, at a median of 1.30 hours. Among those completed within twelve hours, a longer time to completion was associated with higher risk-adjusted in-hospital mortality at an odds ratio of 1.04 per hour, 95 percent confidence interval 1.02 to 1.05, as was a longer time to antibiotics; time to completion of the fluid bolus was not. Seymour CW, Gesten F, Prescott HC, et al. Time to Treatment and Mortality during Mandated Emergency Care for Sepsis. New England Journal of Medicine, 2017;376(23):2235-2244.
    (opens in a new tab)
  8. India performed 8,344,824 cataract operations in 2022-23, against a national target of 7,500,000.The government reported 8,344,824 cataract operations performed nationally in the financial year 2022-23 against a target of 7,500,000. It is the largest single surgical programme the country runs. The operations are done as day care, so ophthalmology ranks far higher by case count than by revenue. Ministry of Health and Family Welfare, Government of India. Cataract surgery performance under the National Programme for Control of Blindness and Visual Impairment. Press Information Bureau, 2023.
    (opens in a new tab)
  9. The national dialysis programme reports 1,889 centres across 759 districts, 262,444 patients, and more than 14.4 million haemodialysis sessions delivered.The programme dashboard reports 1,889 dialysis centres operating across 759 districts, 262,444 patients treated and more than 14.4 million haemodialysis sessions delivered. A session repeats two or three times a week for the life of the patient and is billed at a scheme package rate, so nephrology carries transaction volume out of proportion to its share of revenue. Ministry of Health and Family Welfare, Government of India. Pradhan Mantri National Dialysis Programme, programme dashboard.
    (opens in a new tab)
  10. India's national auditor found 749,820 scheme beneficiaries registered against one mobile number, and 78,396 instances of a single patient recorded as admitted to more than one hospital at the same time.A performance audit of the national health assurance scheme. It recorded 749,820 beneficiaries registered against the mobile number 9999999999, 3,903 claims worth 6.97 crore rupees paid for fresh treatment of 3,446 patients already recorded as dead, and 78,396 instances across 2,231 hospitals of one patient shown as admitted to more than one institution at once. Each of these is a failure of identity and record keeping. None is a finding about clinical care. Comptroller and Auditor General of India. Report No. 11 of 2023: Performance Audit of Ayushman Bharat Pradhan Mantri Jan Arogya Yojana. New Delhi, 2023.
    (opens in a new tab)
  11. In a 2025 study of the national and state assurance schemes in Maharashtra, settlement ran from three months to a year against a thirty day guideline, with individual hospitals carrying two to three crore rupees pending.A mixed methods study of the national scheme and the state scheme in Maharashtra. Claim settlement took three months to a year against a fifteen day intra-state and thirty day portability guideline, individual hospitals carried two to three crore rupees pending, pre-authorisation responses breached the 24 hour guideline, and claims stalled at the account verification stage. Only 796 of roughly 6,000 private hospitals in the state took part at all. Marathe S, Hunter BM, Chakravarthi I, et al. Implementation of publicly funded health insurance schemes in Maharashtra, India. BMC Health Services Research, 2025;25:389.
    (opens in a new tab)
  12. In a prospective study of errors in a hospital stat laboratory, 68.2 percent arose before the sample reached the analyser and a further 18.5 percent after the result left it.Three months of prospective error recording in a hospital stat laboratory. Of the errors found, 68.2 percent were pre-analytical, 13.3 percent analytical and 18.5 percent post-analytical. The dominant categories were specimen and request handling: wrong identification, wrong container, and results transcribed or reported against the wrong request. Plebani M, Carraro P. Mistakes in a stat laboratory: types and frequency. Clinical Chemistry, 1997;43(8):1348-1351.
    (opens in a new tab)
  13. A review of radiological error puts the day to day discrepancy rate between reporting radiologists at 3 to 5 percent, and treats failure to communicate a finding as its own category of harm.A review of the literature on error in diagnostic imaging. It reports a real world day to day discrepancy rate of 3 to 5 percent between reporting radiologists, notes that retrospective rates in study sets containing abnormalities run far higher, and classes the failure to get a finding to the person who can act on it as a category of error of its own, alongside misreading. Brady AP. Error and discrepancy in radiology: inevitable or avoidable? Insights into Imaging, 2017;8(1):171-182.
    (opens in a new tab)
  14. The World Health Organization puts the global cost of medication error at US$42 billion a year and names prescribing, dispensing, administration and monitoring as the four points at which it enters.The third WHO global patient safety challenge. It puts the global cost of medication error at US$42 billion annually, roughly 1 percent of total health expenditure, and identifies prescribing, dispensing, administration and monitoring as the four points where the error is introduced. World Health Organization. Medication Without Harm: WHO Global Patient Safety Challenge. Geneva, 2017.
    (opens in a new tab)
  15. A systematic review found health care associated infection affecting 15.5 patients per 100 admitted in developing countries, against 7.6 per 100 in Europe.A systematic review and meta-analysis of studies from developing countries. Pooled prevalence of health care associated infection was 15.5 per 100 patients, against 7.6 per 100 in Europe and 4.5 per 100 in the United States. Intensive care incidence density was roughly three times the figure reported in high income settings, and surgical site infection was the most frequent category. Allegranzi B, Bagheri Nejad S, Combescure C, et al. Burden of endemic health-care-associated infection in developing countries: systematic review and meta-analysis. The Lancet, 2011;377(9761):228-241.
    (opens in a new tab)
  16. Across 23 acute care hospitals, 48 percent of the surfaces designated for cleaning between patients had in fact been cleaned.An objective evaluation using a transparent marker placed on high touch surfaces before terminal cleaning. Across 23 hospitals, 48 percent of the objects designated for cleaning had been cleaned. Performance varied widely between hospitals and between object types, and feeding the measurement back to the cleaning teams improved it. Carling PC, Parry MF, Von Beheren SM, et al. Identifying opportunities to enhance environmental cleaning in 23 acute care hospitals. Infection Control and Hospital Epidemiology, 2008;29(1):1-7.
    (opens in a new tab)
  17. A review of hospital malnutrition puts its prevalence at 20 to 50 percent of patients on admission and reports that it frequently goes unscreened and unrecorded.A review of prevalence, identification and consequence. Malnutrition is reported in 20 to 50 percent of patients at admission, is frequently neither screened for nor documented, and is associated with longer stay, higher complication rates and higher cost. Barker LA, Gout BS, Crowe TC. Hospital Malnutrition: Prevalence, Identification and Impact on Patients and the Healthcare System. International Journal of Environmental Research and Public Health, 2011;8(2):514-527.
    (opens in a new tab)
  18. WHO guidance on medical equipment maintenance treats a current and complete inventory as the precondition for the programme, because inspection, preventive maintenance, staffing and budget are all sized from it.Guidance on establishing and running a maintenance programme. It treats the equipment inventory as the foundation of the programme, covering inspection, preventive maintenance and corrective maintenance, and states that the inventory determines the staffing, the budget and the measurable performance of the programme. World Health Organization. Medical equipment maintenance programme overview. WHO Medical device technical series, Geneva, 2011.
    (opens in a new tab)
  19. A systematic review associated emergency department crowding with higher inpatient mortality, longer time to treatment for time critical conditions, and more patients leaving before being seen.A systematic review of studies measuring crowding against clinical outcome. Crowding was associated with increased inpatient mortality, longer time to antibiotics in pneumonia and to analgesia in acute pain, longer overall length of stay, and a higher rate of patients leaving without being seen. Bernstein SL, Aronsky D, Duseja R, et al. The effect of emergency department crowding on clinically oriented outcomes. Academic Emergency Medicine, 2009;16(1):1-10.
    (opens in a new tab)
  20. A review of neonatal intensive care found medication errors more frequent than in adult inpatient settings, with dosing the largest single category.A review of published incidence and error type in neonatal intensive care. Reported error rates exceed those of adult inpatient settings, dosing is the most frequent category, and the review attributes the difference to weight based calculation, dilution steps and the narrow margin between a therapeutic and a toxic dose in a neonate. Chedoe I, Molendijk HA, Dittrich STAM, et al. Incidence and nature of medication errors in neonatal intensive care with strategies to improve safety. Drug Safety, 2007;30(6):503-513.
    (opens in a new tab)
  21. A review of monitor alarms reports that between 72 and 99 percent are false, and treats the desensitisation that follows as a patient safety hazard in its own right.A review of the evidence on physiological monitor alarms. Between 72 and 99 percent of alarms are false, staff become desensitised, and alarms are silenced, disabled or set outside safe limits in response. The review makes alarm management the responsibility of the organisation as a whole. Sendelbach S, Funk M. Alarm Fatigue: A Patient Safety Concern. AACN Advanced Critical Care, 2013;24(4):378-386.
    (opens in a new tab)
  22. The United Kingdom's haemovigilance scheme reports that most transfusion incidents are errors in the process, not reactions to the component, and that they cluster at the points where the patient is identified.A review of the first years of the United Kingdom haemovigilance scheme. Most reported incidents were errors, and they clustered in sample collection, laboratory handling and administration at the bedside. At each of those points, the review names patient identification as the recurring failure. Bolton-Maggs PHB, Cohen H. Serious Hazards of Transfusion (SHOT) haemovigilance and progress is improving transfusion safety. British Journal of Haematology, 2013;163(3):303-314.
    (opens in a new tab)
  23. India's national health authority pays an approved hospital 10 percent above the base package rate while it holds entry level accreditation, and 15 percent while it holds full accreditation.The programme document sets an incentive on the base package rate of the national assurance scheme for an empanelled hospital that holds a quality certificate: bronze 5 percent, silver 10 percent, gold 15 percent, with entry level certification valid for two years. Several states have made entry level certification a condition of empanelment. These percentages are the payment rule as published. No outcome was measured to arrive at them. National Accreditation Board for Hospitals and Healthcare Providers. AB-PMJAY Quality Certification Programme. Quality Health Care Programme documentation, accessed 2026.
    (opens in a new tab)
  24. India's National Health Accounts put household out-of-pocket spending at 39.4 percent of total health expenditure in 2021-22, down from 64.2 percent in 2013-14.The national estimates put total health expenditure at 9,04,461 crore rupees, 3.83 percent of gross domestic product and 6,602 rupees per capita. Out-of-pocket expenditure fell to 39.4 percent of total health expenditure from 64.2 percent in 2013-14, while government health expenditure rose from 29 percent to 48 percent of the total. The share is a national account, and it says nothing about any one hospital. Ministry of Health and Family Welfare, Government of India. National Health Accounts Estimates for India 2021-22. Released September 2024.
    (opens in a new tab)
  25. Under India's Digital Health Incentive Scheme a registered facility earns 20 rupees for each digitised record it links, above a threshold of 100 transactions a month.The policy pays a health facility registered on the Health Facility Registry 20 rupees per digitised record created and linked above a threshold of 100 transactions per month, capped per facility, and separately pays 500 rupees or 10 percent of the claim value, whichever is lower, for an insurance claim transacted through the national claims exchange. Both figures are rates the policy sets. They describe the offer, and nothing that followed from it. National Health Authority, Government of India. Financial Incentive Policy under the Digital Health Incentive Scheme, Ayushman Bharat Digital Mission.
    (opens in a new tab)
  26. Across 300 hospitals in nine European countries, each additional patient added to a nurse’s workload was associated with a 7 percent increase in the odds of an inpatient dying within thirty days of admission.A retrospective observational study of 422,730 surgical patients in 300 hospitals across nine countries. Each one patient increase in a nurse’s workload raised the odds of an inpatient dying within thirty days by 7 percent, and every 10 percent rise in the share of nurses holding a bachelor’s degree lowered them by 7 percent. Aiken LH, Sloane DM, Bruyneel L, et al. Nurse staffing and education and hospital mortality in nine European countries: a retrospective observational study. The Lancet, 2014;383(9931):1824-1830.
    (opens in a new tab)
  27. In a registry of more than 370,000 stroke arrivals, patients whose ambulance crew notified the hospital before arrival were treated measurably faster at every step from door to needle.An analysis of 371,988 patients arriving by ambulance at 1,585 hospitals in a national stroke registry. Pre-arrival notification by the ambulance crew was associated with faster door to imaging, faster door to needle, and a higher proportion of eligible patients treated within the time window. Notification happened for only about two thirds of arrivals. Lin CB, Peterson ED, Smith EE, et al. Emergency medical service hospital prenotification is associated with improved evaluation and treatment of acute ischemic stroke. Circulation: Cardiovascular Quality and Outcomes, 2012;5(4):514-522.
    (opens in a new tab)
  28. In a randomised trial, starting physical and occupational therapy while patients were still on the ventilator returned 59 percent of them to independent function at discharge against 35 percent under usual care.A randomised controlled trial in two intensive care units. Therapy begun during the earliest days of mechanical ventilation returned 59 percent of patients to independent functional status at hospital discharge against 35 percent in the control group, with shorter delirium duration and more ventilator free days. Schweickert WD, Pohlman MC, Pohlman AS, et al. Early physical and occupational therapy in mechanically ventilated, critically ill patients: a randomised controlled trial. The Lancet, 2009;373(9678):1874-1882.
    (opens in a new tab)
  29. A pharmacy store audit at an Indian national institute found that roughly 70 percent of the annual drug budget sat in about 10 percent of the items on the shelf.An inventory analysis of a teaching institute pharmacy store combining annual consumption value with clinical criticality. A small minority of items carried most of the annual expenditure, while a separate and only partly overlapping set carried the items whose absence stops treatment. The study names stockout, expiry and supplier failure as the recurring failures, and traces overstocking to habit, not demand. Kumar S, Chakravarty A, et al. ABC-VED analysis of the medical store of a national institute in eastern India. Journal of Family Medicine and Primary Care, 2023.
    (opens in a new tab)
  30. India had about 2,150 accredited medical testing laboratories as of February 2024.The national accreditation board publishes a running count of laboratories holding its accreditation. The count comes from the board itself. Quality in any single laboratory is outside what it measures. National Accreditation Board for Testing and Calibration Laboratories. Number of accredited laboratories, accessed 2026.
    (opens in a new tab)
  31. A national study of hospitals performing primary angioplasty identified six specific process changes, among them activation of the catheterisation laboratory by the emergency physician, each independently associated with a significantly shorter door to balloon time.A cross sectional study of 365 hospitals performing primary percutaneous coronary intervention. Six strategies were independently associated with faster treatment. Each changed how the process was organised and left the procedure itself alone: activation of the laboratory by the emergency physician, activation by a single call to a central page operator, activation while the patient was still en route, an expectation that staff arrive within twenty minutes of being paged, an attending cardiologist always on site, and real time feedback of times to the staff involved. Bradley EH, Herrin J, Wang Y, et al. Strategies for Reducing the Door-to-Balloon Time in Acute Myocardial Infarction. New England Journal of Medicine, 2006;355(22):2308-2320.
    (opens in a new tab)
  32. Median door to balloon time in the United States fell from 96 minutes in 2005 to 64 minutes in 2010, and the share of patients treated within 90 minutes rose from 44.2 percent to 91.4 percent.An analysis of national registry data across five years of hospital reporting. Median door to balloon time fell from 96 minutes to 64 minutes and the proportion treated inside the 90 minute standard rose from 44.2 percent to 91.4 percent. The improvement was spread across hospitals of every size. Krumholz HM, Herrin J, Miller LE, et al. Improvements in Door-to-Balloon Time in the United States, 2005 to 2010. Circulation, 2011;124(9):1038-1045.
    (opens in a new tab)
  33. Across primary angioplasty admissions at 423 hospitals, a shorter door to balloon time was associated with lower in hospital mortality within each year studied.A retrospective study of primary percutaneous coronary intervention admissions. Within any given year, patients treated faster had lower in hospital mortality; across the years, the national fall in treatment times was not accompanied by a corresponding fall in annual mortality, which the authors attribute in part to a changing case mix. The paper is cited here for the within year association between delay and death, which is what a clock inside a hospital can act on. Nallamothu BK, Normand SL, Wang Y, et al. Relation between door-to-balloon times and mortality after primary percutaneous coronary intervention over time: a retrospective study. The Lancet, 2015;385(9973):1114-1122.
    (opens in a new tab)
  34. Across 19 hospitals running an automated deterioration alert with a defined clinical response, patients who reached the alert threshold had lower mortality within 30 days than comparable patients before deployment, at an adjusted relative risk of 0.84.A deployment study across an integrated hospital system. A predictive score ran continuously against the record, and its alerts went to a defined responder. Among patients who met the alert threshold, mortality within 30 days and intensive care admission after the alert were both lower than in the comparison period. The authors are explicit that the study evaluated the score and the response protocol together. Escobar GJ, Liu VX, Schuler A, et al. Automated Identification of Adults at Risk for In-Hospital Clinical Deterioration. New England Journal of Medicine, 2020;383(20):1951-1960.
    (opens in a new tab)
  35. Caesarean section use nearly doubled worldwide between 2000 and 2015, from 12.1 percent to 21.1 percent of births, and the widest disparities lay within countries.An analysis of national data for 169 countries. Global caesarean use rose from 12.1 percent of births in 2000 to 21.1 percent in 2015. Within countries the gap between the richest and poorest fifths, and between private and public facilities, was consistently larger than the gap between countries at similar income levels. Boerma T, Ronsmans C, Melesse DY, et al. Global epidemiology of use of and disparities in caesarean sections. The Lancet, 2018;392(10155):1341-1348.
    (opens in a new tab)
  36. Hospitals in a state perinatal quality collaborative that adopted a standard obstetric haemorrhage protocol reduced severe maternal morbidity among haemorrhage patients by 20.8 percent, against 1.2 percent at comparison hospitals over the same period.A collaborative across 99 hospitals implementing a standard haemorrhage toolkit: a risk assessment on admission, quantitative measurement of blood loss, a stocked haemorrhage cart, and a staged response protocol. Severe maternal morbidity among women with haemorrhage fell 20.8 percent in participating hospitals against 1.2 percent in comparison hospitals. Main EK, Cape V, Abreo A, et al. Reduction of severe maternal morbidity from hemorrhage using a state perinatal quality collaborative. American Journal of Obstetrics and Gynecology, 2017;216(3):298.e1-298.e11.
    (opens in a new tab)
  37. In an electronic record that raised an automated alert on an abnormal imaging result, 7.7 percent of alerts were never acknowledged and 4.0 percent of the abnormal results still lacked timely follow up four weeks later.A prospective study of automated alerts for abnormal imaging in an outpatient setting with a mature electronic record. Alerts that reached nobody in particular were the failure mode: unacknowledged alerts and alerts sent to more than one recipient were both associated with a lack of timely follow up. The record delivered the result; ownership of it was what the system did not carry. Singh H, Thomas EJ, Mani S, et al. Timely follow-up of abnormal diagnostic imaging test results in an outpatient setting: are electronic medical records achieving their potential? Archives of Internal Medicine, 2009;169(17):1578-1586.
    (opens in a new tab)
  38. Diagnostic error was estimated to affect about 5.08 percent of United States adults each year, approximately 12 million people, with about half of those errors carrying the potential for severe harm.An estimate built from three independent observational studies using electronic triggers to find missed and delayed diagnoses in outpatient care. The pooled rate was 5.08 percent of adults a year, about 12 million people, and roughly half of the errors identified had the potential to cause severe harm. Singh H, Meyer AND, Thomas EJ. The frequency of diagnostic errors in outpatient care: estimations from three large observational studies involving US adult populations. BMJ Quality and Safety, 2014;23(9):727-731.
    (opens in a new tab)
  39. In a randomised trial of routine symptom reporting during chemotherapy, median overall survival was 31.2 months with monitoring against 26.0 months under usual care.A randomised trial at a single centre in which patients receiving chemotherapy reported symptoms through a structured form, with an alert to the nursing team when a symptom worsened. Median overall survival was 31.2 months in the reporting group against 26.0 months in the usual care group. The intervention was a form, a route for it and a responder. Neither group received a new therapy. Basch E, Deal AM, Dueck AC, et al. Overall Survival Results of a Trial Assessing Patient-Reported Outcomes for Symptom Monitoring During Routine Cancer Treatment. JAMA, 2017;318(2):197-198.
    (opens in a new tab)
  40. After multidisciplinary team working was introduced across a health region, breast cancer mortality at five years was 18 percent lower than in comparison regions, across 13,722 women.An interrupted time series comparing one health region that introduced mandatory multidisciplinary team review against neighbouring regions that had not yet done so. Breast cancer mortality at five years was 18 percent lower in the intervention region after introduction, and the difference was not present before it. Kesson EM, Allardice GM, George WD, et al. Effects of multidisciplinary team working on breast cancer survival: retrospective, comparative, interventional cohort study of 13 722 women. BMJ, 2012;344:e2718.
    (opens in a new tab)
  41. Across fourteen cardiovascular and cancer operations, adjusted mortality was lower at higher volume hospitals for every one of them, with the absolute difference running from just over 0.2 percent to more than 12 percent depending on the procedure.A national analysis of six cardiovascular procedures and eight major cancer resections. Adjusted mortality fell as hospital volume rose for all fourteen. The absolute difference between very low and very high volume hospitals was smallest for carotid endarterectomy and largest for pancreatic resection, where it exceeded 12 percentage points. Birkmeyer JD, Siewers AE, Finlayson EVA, et al. Hospital Volume and Surgical Mortality in the United States. New England Journal of Medicine, 2002;346(15):1128-1137.
    (opens in a new tab)
  42. The Global Task Force on Radiotherapy for Cancer Control estimated that scaling up radiotherapy capacity in low and middle income countries by 2035 would save 26.9 million life years over the lifetimes of the patients treated.A modelling study for the Global Task Force on Radiotherapy for Cancer Control, convened by the Lancet Oncology Commission. It estimated the capacity gap, the investment required to close it, and the health and economic return. Scaling up capacity in low and middle income countries by 2035 was estimated to save 26.9 million life years, with returns exceeding the investment on every scenario the authors modelled. Atun R, Jaffray DA, Barton MB, et al. Expanding global access to radiotherapy. The Lancet Oncology, 2015;16(10):1153-1186.
    (opens in a new tab)
  43. A study that quantified individual quality control checks in radiation oncology found wide variation in how many errors each check caught, with physics review of the plan and the chart among the most effective.A prospective study in one radiation oncology department that recorded which quality control check detected each error and how far the error had travelled before detection. Detection was unevenly distributed across the checks, and physics review of the plan and the chart accounted for a large share of it. The authors argue that a department can only allocate quality control effort once it measures which check is doing the work. Ford EC, Terezakis S, Souranis A, et al. Quality control quantification (QCQ): a tool to measure the value of quality control checks in radiation oncology. International Journal of Radiation Oncology, Biology, Physics, 2012;84(3):e263-e269.
    (opens in a new tab)
  44. Among patients sent home from an emergency department, the risk of death or admission within seven days rose steadily with the mean length of stay of the department they were seen in.A population based cohort study of nearly 14 million emergency department visits over five years. For patients discharged home, the adjusted odds of death and of admission within seven days rose as the mean length of stay of the department increased, and the association held after adjustment for the patient and for the hospital. Guttmann A, Schull MJ, Vermeulen MJ, et al. Association between waiting times and short term mortality and hospital admission after departure from emergency department: population based cohort study from Ontario, Canada. BMJ, 2011;342:d2983.
    (opens in a new tab)
  45. In an external validation at one academic health system, a widely implemented proprietary sepsis prediction model discriminated at an area under the curve of 0.63 and identified 33 percent of the patients who had sepsis.A retrospective cohort of 27,697 patients across 38,455 hospitalisations at one academic health system, evaluating a model that was in use at hundreds of hospitals. Area under the curve was 0.63, sensitivity 33 percent, specificity 83 percent, and positive predictive value 12 percent at the recommended threshold. The model raised an alert on 18 percent of all hospitalised patients and did not identify 1,709 of the 2,552 patients who had sepsis. Wong A, Otles E, Donnelly JP, et al. External Validation of a Widely Implemented Proprietary Sepsis Prediction Model in Hospitalized Patients. JAMA Internal Medicine, 2021;181(8):1065-1070.
    (opens in a new tab)
  46. An algorithm applied to roughly 200 million people a year used health cost as a proxy for health need, and changing what it was asked to predict raised the share of Black patients identified for additional care from 17.7 percent to 46.5 percent.An analysis of a commercial risk prediction tool applied at national scale to decide which patients received additional care. At any given risk score, Black patients were measurably sicker than white patients, because the model had been trained to predict cost. Illness was never its target. The authors conclude that the choice of a convenient proxy for ground truth is itself a source of algorithmic bias, and that the remedy was to change the label. Obermeyer Z, Powers B, Vogeli C, Mullainathan S. Dissecting racial bias in an algorithm used to manage the health of populations. Science, 2019;366(6464):447-453.
    (opens in a new tab)
  47. Across eleven clinics, a screening system whose reported laboratory accuracy was above 90 percent rejected 21 percent of the images submitted to it as too low in quality to process.A field study of a deployed screening system across eleven clinics over roughly eight months, on images captured by nurses in ordinary clinic conditions. Around one image in five was rejected before it reached the model, largely because of clinic lighting, and slow network links held throughput to about ten patients per two hour session. The failures the study reports came from the clinics and from how the images were captured, not from the model. Beede E, Baylor E, Hersch F, et al. A Human-Centered Evaluation of a Deep Learning System Deployed in Clinics for the Detection of Diabetic Retinopathy. Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems.
    (opens in a new tab)
  48. Of 521 regulator-authorised artificial intelligence enabled medical devices reviewed, 226 had no published clinical validation data.A review of 521 devices authorised between 1995 and 2022. Roughly 43 percent had no published clinical validation data of any kind, and only about a quarter reported prospective validation. Annual authorisations rose from 2 in 2016 to 69 in the last year studied, and most were cleared by comparison with an existing device. A fresh clinical evidence submission was the less common route. Muralidharan V, Ng MY, Adewale BA, et al. Not all AI health tools with regulatory authorization are clinically validated. Nature Medicine, 2024.
    (opens in a new tab)
  49. Among United States hospitals running predictive models inside the electronic record, 61 percent evaluated most or all of those models locally for accuracy and 44 percent evaluated them for bias.A national survey of hospital information technology adoption. 65 percent of hospitals used artificial intelligence or predictive models integrated with the record. Of those, 79 percent took models from their record vendor. Local evaluation for accuracy and for bias was concentrated in hospitals with higher operating margins and system membership, which the authors read as a warning that hospitals serving marginalised populations may not be able to check what they deploy. Nong P, Adler-Milstein J, Apathy NC, Holmgren AJ, Everson J. Current Use And Evaluation Of Artificial Intelligence And Predictive Models In US Hospitals. Health Affairs, 2025;44(1):90-98.
    (opens in a new tab)
  50. Among 53 practitioners building high-stakes machine learning systems in India, East and West Africa and the United States, 92 percent reported at least one data cascade: a failure that surfaces downstream but starts in the data.A semi-structured interview study of practitioners working in high-stakes domains including cancer detection. Data cascades are described as compounding events that begin upstream in collection, definition or documentation and surface late, as unexplained model behaviour, where they are expensive to diagnose and are commonly misread as model problems. The study reports that the work of defining and maintaining data is systematically undervalued relative to the work of building models. Sambasivan N, Kapania S, Highfill H, Akrong D, Paritosh P, Aroyo LM. Everyone wants to do the model work, not the data work: Data Cascades in High-Stakes AI. Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems.
    (opens in a new tab)
  51. A clinical review names dataset shift, the divergence between the data a model was trained on and the data it later meets, as a major driver of artificial intelligence system malfunction, and directs clinicians to establish monitoring and updating protocols.A review written for clinicians. It enumerates the ordinary hospital events that move a model away from the population it was fitted to, among them a change of supplier, a change of coding practice, a new unit, a new patient mix and a change in how a variable is recorded. Its instruction to the deploying institution is that a model in service needs monitoring and a defined updating protocol. A performance figure quoted at purchase is not enough. Finlayson SG, Subbaswamy A, Singh K, et al. The Clinician and Dataset Shift in Artificial Intelligence. New England Journal of Medicine, 2021;385(3):283-286.
    (opens in a new tab)
  52. An independent patient safety organisation ranked risks with artificial intelligence enabled health technologies first on its 2025 list of the top ten health technology hazards.An annual hazard list published by an independent non-profit patient safety organisation, drawn from its own device testing, incident investigations and reporting databases. For 2025 the first-ranked hazard was risks with artificial intelligence enabled health technologies, placed above cybersecurity attacks and above substandard or falsified medical products. The list ranks hazards for hospitals to manage. It makes no count of harm that has already occurred. ECRI. Top 10 Health Technology Hazards for 2025. ECRI, 2025.
    (opens in a new tab)
  53. In a sixteen country survey, 76 percent of Indian healthcare professionals said artificial intelligence could improve patient outcomes, and 40 percent said the tools available to them fit their day-to-day clinical needs.A survey of more than 1,900 healthcare professionals and more than 16,000 patients across sixteen countries, fielded between December 2024 and April 2025 and published by a medical technology manufacturer, without peer review. In the India findings, 76 percent of professionals were optimistic that artificial intelligence can improve patient outcomes and 80 percent expected it to reduce administrative burden, while 40 percent said current tools align with their day-to-day clinical needs. 45 percent asked for clear guidance on use and limitations and 44 percent for clarity on legal liability. It is cited here for the gap between two of its own figures: the 76 percent who are optimistic, and the 40 percent who say today's tools fit their day-to-day clinical needs. Philips. Future Health Index 2025: Building Trust in Healthcare AI. India findings, 2025.
    (opens in a new tab)

Powerful. Responsible. Pensieve.

Get started