"Our clinicians stopped abandoning literature searches mid-shift — the answer arrives with its citation, in the time it takes to ask the question."
Clinicians were skipping evidence checks mid-shift because the search took longer than the patient encounter itself. The team built a clinical Q&A service that pulls answers directly from PubMed and the patient's own chart — and every answer carries the exact passage it came from. Responses come back in under a second with citations attached, and the audit trail is automatic. Senior physicians stopped second-guessing the tool the day they could trace every answer back to its source.
"Phone lines that used to ring through the morning now route themselves — and the staff we freed up went straight to the bedside."
The front desk was drowning. Bookings, reschedules, urgent-care callers — all on the same line, all needing a human to answer. The voice agent picks up directly now, books and triages by urgency, and a calm conversational layer handles anxious callers while routing the rest to staff who actually need to be on the call. Scheduling workload dropped 40%, no-shows dropped 25%, and patient satisfaction lifted 15% — and the team that used to be tied to the phone went back to clinical work.
"Decision support stopped being a black box the day every answer started citing the EHR passage it came from."
Decision-support tools were getting pushback from clinicians who couldn't see where the recommendations came from. The team built a system that pulls insights from years of EHR data, and every answer it gives carries the exact passage it was drawn from — so a physician can see what the model saw before acting on it. The pipeline runs on modest hospital hardware and returns answers in under a second. Adoption climbed once clinicians could trace every recommendation back to the chart.
"We caught arrhythmia patterns from a wristband that previously needed a hospital monitor — and we caught them before the patient knew anything was wrong."
Arrhythmias were going undetected between hospital visits because consumer wearables couldn't reliably flag them. The team trained a deep-learning model on the data already streaming from wrist sensors — heart rhythm, motion, orientation — and tuned it to find irregular patterns without flooding the wearer with false alarms. The result was hospital-grade detection from a device the patient was already wearing, with the model catching events earlier and more accurately than the standard baseline.
"We stopped designing one interface for everyone and started letting the interface read the user — eye by eye, click by click."
Standard digital banking interfaces were leaving neurodivergent users behind — autism, ADHD, and dyslexia each demand different layouts, pacing, and cues. The team built an AI that watches how a user actually interacts with the screen, picks up the patterns through eye-tracking and behavior, and quietly reshapes the interface — adjusting content, navigation, calls-to-action, even color. Each session leaves the system a little smarter, so the next time the user opens the app, it already knows what works for them.
"Our case managers stopped reacting to overdose events and started intervening months before — every flag the model raised gave us a window we never had."
Case managers were always one step behind on opioid risk — they saw patients after a crisis, not before. The team built a risk model trained on prescription patterns, demographics, and historical claims, balanced carefully across age, gender, and race so it wouldn't misread any group. The model flagged at-risk patients with high accuracy and — just as importantly — almost never raised a false alarm on low-risk ones. Intervention teams could finally focus on the patients who actually needed them.
"Budget season used to start with a spreadsheet and a guess — now it starts with a forecast every department head can defend."
Workforce planning was an annual exercise in educated guesswork — leadership knew demand was changing, but had no way to model what staffing they'd need three or five years out. The team built a forecasting engine that projects patient demand across multiple horizons, translates that into department-level workload using standard healthcare staffing ratios, and predicts turnover by specialty. Department heads now bring forecasts into budget conversations instead of intuition, and gap analysis flags shortfalls before they hit the floor.
"Reads that used to wait days for senior review now come pre-stratified — radiologists see the equivocal ones first, and the noise gets quieter."
Radiology read queues were piling up, and DICOM scans from different vendors made consistent triage harder. The team built a model that pre-reads every scan as it arrives, flags the positive cases, and stratifies the queue so radiologists see the equivocal scans first. The preprocessing pipeline handles the vendor variation automatically — no per-site retraining — making the model deployment-ready across mixed imaging estates from day one.
"Trial enrollment used to bottleneck on a coordinator reading every chart — now the chart reads itself, and the coordinator approves."
Clinical trial enrollment was bottlenecked on coordinators reading patient charts one by one to check eligibility — slow, expensive, and inconsistent. The team built a model that reads clinical notes directly, scores eligibility against trial criteria, and flags the right candidates without human triage. Coordinators went from chart-readers to decision-makers — reviewing only the borderline cases — and oncology trials filled their cohorts faster than they had in years.
"Patient-level predictors we used to extract by hand now lift out of the notes themselves — and the SVM sees what the chart reviewer might miss."
NICU teams were spending hours abstracting jaundice indicators from free-text clinical notes — discharge summaries, nursing notes, radiology, physician progress reports — each in different writing styles. The team built a classifier that reads the notes directly and identifies jaundice with near-perfect precision. The cross-functional build — data science, clinical informatics, and BI engineers in the same room — meant the model was usable from day one, and the chart reviewers got back to work that needed their judgment.
"Our care plan stopped being a snapshot at discharge and became a living signal — every gait, every fall risk, every recovery trajectory."
Once a patient left the hospital, the care team lost visibility — recovery progress was a guess until the next visit. The team trained a deep-learning model on the data already coming off patient wearables — motion, orientation, heart rhythm — and taught it to classify activity continuously. The model now flags drops in mobility, gait changes, and recovery plateaus in near real-time, so clinical teams catch setbacks while there's still time to act.
"Our fraud team stopped chasing the easy alerts and started seeing the patterns nobody had named — the kind that don't make it into rule books."
Payer fraud teams were chasing alerts from rule-based systems that missed every novel scheme. The team built a dual-model pipeline: one model scores patients for segmentation and resource allocation, the other classifies fraud, waste, and abuse across the full claims data — beneficiary, inpatient, outpatient — at near-perfect accuracy. The SIU team finally had a system that found patterns instead of just confirming the ones already in the policy manual.
"Women who used to wait months for a diagnosis now self-screen in minutes — and reach the operating room only when the data says they should."
Uterine fibroid diagnosis was taking three to six months — symptoms went unreported, primary care missed the early signals, and women reached specialist care only after the condition was advanced. The team built a mobile app that captures symptoms and risk factors directly from the patient, scores fibroid risk in real time against clinical baselines, and routes the right cases into teleconsultation — with insurance and compliance handled inline. Diagnosis time collapsed from months to under a week, and nearly half the cases that would have gone to surgery were resolved without it.
"We stopped asking 'who fell last month' and started knowing 'who will fall this month' — the survey became targeted, the prevention became real."
Fall risk among seniors was being assessed reactively — surveys went out to everyone, falls happened anyway, and the cost (both human and financial) kept climbing. The team built a predictive model fusing claims, survey responses, and demographics to find the specific drug interactions and comorbidity patterns that precede falls. Outreach went from blanket coverage to targeted intervention, and the model correctly identified almost every actual fall patient before the event.
"The radiologist's eye still calls the verdict — but the model finds the lesion before the eye gets there."
Radiologists were reading chest X-rays at volumes that left no room for second-pass scrutiny — the model needed to do the looking that the human couldn't. The team trained a segmentation model to mark cancerous lung regions pixel by pixel, validated it iteratively against held-out test data, and integrated the predicted mask directly into the radiologist's workflow. The radiologist still makes the call — but now they're calling on a pre-marked image, not a blank one.
"Allocating savings used to be a quarterly negotiation — now it's a daily dashboard, and every initiative knows what it earned."
A national health insurer was running multiple cost-control initiatives simultaneously, with no fair way to attribute savings to each one. Budget season turned into a political negotiation. The team built an end-to-end framework — three years of claims, case management, and prior authorization data feeding budget models, savings estimation, FWA detection, and a fair-allocation engine. Each initiative now sees its actual contribution, in production, daily — and the dashboards work for C-suite, ops, and regulators alike.
"Refiling claims used to be a department — now it's a dashboard, and the denials we can avoid never get filed in the first place."
Claims refiling was eating headcount — a large team spent days reworking denials that were predictable from the moment the claim was generated. The team built a model that scores every claim on its way out the door and surfaces the underlying drivers. Primary payer and policy type turned out to be the dominant denial signals — actionable intelligence that fed renegotiation strategies and turned refile from a permanent department into a dashboard.
"Underpayment used to hide inside the aggregate — now every provider sees which payer underpaid which CPT code, and acts on it."
Providers knew they were being underpaid — they just couldn't see where. The aggregate numbers looked fine; the leakage was distributed across hundreds of payer-code combinations. The team built a clustering engine that mines the claim payment data across multiple dimensions and surfaces underpayments per payer and per denial reason — measured in denied dollars per invoice and ready for direct negotiation.
"The chart finally reads itself the way a physician would — into the right pathway, the right category, the right next step."
Clinical notes were piling up in inboxes that nobody had time to sort — every chart had to be read and triaged into a pathway by hand. The team fine-tuned a language model to read notes and route them into five disease categories with strong accuracy, freeing physicians to act on the chart instead of sorting it. The model is interpretable enough to debug, stable enough to deploy at scale, and reduces routing latency across the entire diagnostic queue.
"Discharge planning stopped beginning at day five — it now begins when the patient walks through the door."
Diabetic inpatient stays were stretching well past clinical necessity — and the care team only knew which patients would stay long after they already had. The team trained a neural network to flag high-LOS patients at the moment of admission, using demographics, comorbidities, and lab values. A parallel regression identifies the specific comorbidities driving extended stays — so the care team can plan discharge, intervene early, and forecast bed utilization before complications cascade.
"Our pediatric dentists stopped squinting at small mouths on small films — the model points where to look, and the chair time gets returned to the child."
Pediatric dental X-rays are hard — small mouths, restless patients, films that don't sit still. Dentists were spending more time finding the lesion than treating it. The team trained a detection model to identify caries and apical periodontitis automatically and overlay color-coded bounding boxes on the X-ray. The model points the dentist's eye to the lesion immediately — returning chair time to the patient and reducing the cognitive load of every read.
"We finally had a CHF readmission model that beat the public benchmarks — and the readmissions we prevented paid for the program in the first year."
CHF readmissions were the costliest preventable problem on the inpatient floor — and the standard CMS risk models weren't accurate enough to justify intervention spend. The team built a custom readmission model trained on 12 months of pre-admission history and 30 days of post-discharge data, and it outperformed both the CMS HF and CMS AMI benchmarks. The intervention engine layered on top now feeds tailored recommendations to discharge planners — and the readmissions prevented paid back the program in year one.
"Nephrologists got a real-time map of which patients were sliding toward dialysis — and intervened where the slide could still be slowed."
CKD Stage 3 patients progress toward dialysis at very different rates — and nephrologists were having to read every chart by hand to figure out who needed urgent intervention. The team trained a model on three to five years of longitudinal lab data and built dashboards that score patients by progression speed in real time. Nephrologists now see, at a glance, which patients are sliding fast — and intervene where the slide can still be slowed.
"Therapists used to triage by intuition — now the alert tells them which patient is hours away from an avoidable hospitalization."
A large home health and hospice provider was running on therapist intuition — visits prioritized by gut feel, with no early signal for which patients were sliding toward avoidable hospitalization. The team built a four-model platform on a unified Azure research lake — predicting acute hospitalization risk, 90-day mortality, ADL improvement, and mobility outcomes — and pushed the alerts directly to therapists in persona-tuned dashboards. Triage stopped being a guess.
"Pricing used to come from an actuarial table — now it comes from the data the actuary trusted but couldn't fully read."
Insurance charge models were stuck in traditional actuarial tables — useful, but not granular enough to support modern pricing decisions. The team built a regression model that learns charge patterns from individual risk factors — age, BMI, smoking, dependents, region — and surfaces feature importance transparently. Pricing teams now see why a given premium falls where it does, which segments to target, and where the actuarial table was leaving money on the table.
"Treatment planning used to start with a tracing pencil — now it starts with a mask, and the orthodontist gets back the half-hour."
Orthodontic treatment planning was starting with a tracing pencil — manual tooth segmentation on every panoramic X-ray, billable hours spent on a task no orthodontist enjoyed. The team trained a segmentation model to outline each tooth automatically with clinical-grade accuracy. The model integrates straight into the imaging workflow, returning half-hour blocks of time to the clinician — and the segmentation is consistent across the practice, not artist-by-artist.
"Pediatric films are harder, the patients more variable — the model still finds the structure, even when the mouth doesn't sit still."
Pediatric tooth segmentation is harder than adult — developing dentition varies more, films are noisier, and patients move. The team adapted a segmentation model for pediatric X-rays with carefully standardized preprocessing, and it now segments individual teeth reliably even on the difficult films. Orthodontists and pediatric dentists use the output for treatment planning, growth monitoring, and faster diagnostic workflows on the population that needs them most.
"Cross-sell stopped being a phone-list and started being a targeted conversation — fewer calls, higher conversion, less noise for the policyholder."
Insurance cross-sell was being run as a volume play — call every policyholder, hope a fraction convert. Conversion rates were low, customer fatigue was high. The team built a propensity model that scores health-insurance policyholders for vehicle-insurance interest using their existing data — demographic, policy, and vehicle features. Targeted lists replaced bulk calling, conversion climbed, and policyholders stopped getting calls they didn't want.
"Lesion borders used to depend on which dermatologist drew them — now the mask is the same on Monday morning as it is on Friday evening."
Dermatologists were drawing lesion borders by hand — and the borders moved depending on who drew them, when, and how tired they were. The team trained a segmentation model on standardized dermoscopic images with rotation and flip augmentation, and the model now outlines lesions pixel by pixel with consistent precision. Dermatologists use the mask as a baseline for screening — the consistency means earlier detection and fewer missed margins.
"Stroke risk used to live in the GP's instinct — now it lives in a number the patient can see and the population health team can act on."
Stroke risk assessment was being done in the GP's head — different doctor, different read. Population health teams had no consistent signal to organize prevention around. The team trained a classifier on clinical and lifestyle data — age, BMI, smoking, hypertension, glucose — and tuned it carefully to avoid bias toward majority classes. The model now produces a defensible stroke-risk score the patient can see and the population health team can act on at scale.
"Radiology second-reads used to depend on bandwidth — now the model gives every image a second look, and the senior radiologist arbitrates only the equivocal ones."
Senior radiologist second-reads were the gold standard for breast cancer screening — and the bottleneck. Junior reads went unverified, and the senior team couldn't keep up with volume. The team built a classifier that gives every image a second read automatically — benign or malignant — with augmentation and early-stopping tuned for clinical-grade generalization. Senior radiologists now spend their attention only on the equivocal cases, and the easy ones come pre-validated.
"A scan that took two senior radiologists to call now lands on the workstation with a probability and a Hounsfield trace — and the gut check happens faster."
CT classification for lung tumors had been a yes-or-no call — and that wasn't granular enough for treatment planning. The team trained a 3D model to classify scans into three categories — normal, malignant, benign — with preprocessing that normalizes Hounsfield Units and isolates the region of interest before classification. Radiologists now see a probability across three classes instead of a binary flag, and the model converges quickly and stably enough for reliable production support.
"Our clinicians stopped abandoning literature searches mid-shift — the answer arrives with its citation, in the time it takes to ask the question."
Clinicians were skipping evidence checks mid-shift because the search took longer than the patient encounter itself. The team built a clinical Q&A service that pulls answers directly from PubMed and the patient's own chart — and every answer carries the exact passage it came from. Responses come back in under a second with citations attached, and the audit trail is automatic. Senior physicians stopped second-guessing the tool the day they could trace every answer back to its source.
"Phone lines that used to ring through the morning now route themselves — and the staff we freed up went straight to the bedside."
The front desk was drowning. Bookings, reschedules, urgent-care callers — all on the same line, all needing a human to answer. The voice agent picks up directly now, books and triages by urgency, and a calm conversational layer handles anxious callers while routing the rest to staff who actually need to be on the call. Scheduling workload dropped 40%, no-shows dropped 25%, and patient satisfaction lifted 15% — and the team that used to be tied to the phone went back to clinical work.
"Decision support stopped being a black box the day every answer started citing the EHR passage it came from."
Decision-support tools were getting pushback from clinicians who couldn't see where the recommendations came from. The team built a system that pulls insights from years of EHR data, and every answer it gives carries the exact passage it was drawn from — so a physician can see what the model saw before acting on it. The pipeline runs on modest hospital hardware and returns answers in under a second. Adoption climbed once clinicians could trace every recommendation back to the chart.
"We caught arrhythmia patterns from a wristband that previously needed a hospital monitor — and we caught them before the patient knew anything was wrong."
Arrhythmias were going undetected between hospital visits because consumer wearables couldn't reliably flag them. The team trained a deep-learning model on the data already streaming from wrist sensors — heart rhythm, motion, orientation — and tuned it to find irregular patterns without flooding the wearer with false alarms. The result was hospital-grade detection from a device the patient was already wearing, with the model catching events earlier and more accurately than the standard baseline.
"We stopped designing one interface for everyone and started letting the interface read the user — eye by eye, click by click."
Standard digital banking interfaces were leaving neurodivergent users behind — autism, ADHD, and dyslexia each demand different layouts, pacing, and cues. The team built an AI that watches how a user actually interacts with the screen, picks up the patterns through eye-tracking and behavior, and quietly reshapes the interface — adjusting content, navigation, calls-to-action, even color. Each session leaves the system a little smarter, so the next time the user opens the app, it already knows what works for them.
"Our case managers stopped reacting to overdose events and started intervening months before — every flag the model raised gave us a window we never had."
Case managers were always one step behind on opioid risk — they saw patients after a crisis, not before. The team built a risk model trained on prescription patterns, demographics, and historical claims, balanced carefully across age, gender, and race so it wouldn't misread any group. The model flagged at-risk patients with high accuracy and — just as importantly — almost never raised a false alarm on low-risk ones. Intervention teams could finally focus on the patients who actually needed them.
"Budget season used to start with a spreadsheet and a guess — now it starts with a forecast every department head can defend."
Workforce planning was an annual exercise in educated guesswork — leadership knew demand was changing, but had no way to model what staffing they'd need three or five years out. The team built a forecasting engine that projects patient demand across multiple horizons, translates that into department-level workload using standard healthcare staffing ratios, and predicts turnover by specialty. Department heads now bring forecasts into budget conversations instead of intuition, and gap analysis flags shortfalls before they hit the floor.
"Reads that used to wait days for senior review now come pre-stratified — radiologists see the equivocal ones first, and the noise gets quieter."
Radiology read queues were piling up, and DICOM scans from different vendors made consistent triage harder. The team built a model that pre-reads every scan as it arrives, flags the positive cases, and stratifies the queue so radiologists see the equivocal scans first. The preprocessing pipeline handles the vendor variation automatically — no per-site retraining — making the model deployment-ready across mixed imaging estates from day one.
"Trial enrollment used to bottleneck on a coordinator reading every chart — now the chart reads itself, and the coordinator approves."
Clinical trial enrollment was bottlenecked on coordinators reading patient charts one by one to check eligibility — slow, expensive, and inconsistent. The team built a model that reads clinical notes directly, scores eligibility against trial criteria, and flags the right candidates without human triage. Coordinators went from chart-readers to decision-makers — reviewing only the borderline cases — and oncology trials filled their cohorts faster than they had in years.
"Patient-level predictors we used to extract by hand now lift out of the notes themselves — and the SVM sees what the chart reviewer might miss."
NICU teams were spending hours abstracting jaundice indicators from free-text clinical notes — discharge summaries, nursing notes, radiology, physician progress reports — each in different writing styles. The team built a classifier that reads the notes directly and identifies jaundice with near-perfect precision. The cross-functional build — data science, clinical informatics, and BI engineers in the same room — meant the model was usable from day one, and the chart reviewers got back to work that needed their judgment.
"Our care plan stopped being a snapshot at discharge and became a living signal — every gait, every fall risk, every recovery trajectory."
Once a patient left the hospital, the care team lost visibility — recovery progress was a guess until the next visit. The team trained a deep-learning model on the data already coming off patient wearables — motion, orientation, heart rhythm — and taught it to classify activity continuously. The model now flags drops in mobility, gait changes, and recovery plateaus in near real-time, so clinical teams catch setbacks while there's still time to act.
"Our fraud team stopped chasing the easy alerts and started seeing the patterns nobody had named — the kind that don't make it into rule books."
Payer fraud teams were chasing alerts from rule-based systems that missed every novel scheme. The team built a dual-model pipeline: one model scores patients for segmentation and resource allocation, the other classifies fraud, waste, and abuse across the full claims data — beneficiary, inpatient, outpatient — at near-perfect accuracy. The SIU team finally had a system that found patterns instead of just confirming the ones already in the policy manual.
"Women who used to wait months for a diagnosis now self-screen in minutes — and reach the operating room only when the data says they should."
Uterine fibroid diagnosis was taking three to six months — symptoms went unreported, primary care missed the early signals, and women reached specialist care only after the condition was advanced. The team built a mobile app that captures symptoms and risk factors directly from the patient, scores fibroid risk in real time against clinical baselines, and routes the right cases into teleconsultation — with insurance and compliance handled inline. Diagnosis time collapsed from months to under a week, and nearly half the cases that would have gone to surgery were resolved without it.
"We stopped asking 'who fell last month' and started knowing 'who will fall this month' — the survey became targeted, the prevention became real."
Fall risk among seniors was being assessed reactively — surveys went out to everyone, falls happened anyway, and the cost (both human and financial) kept climbing. The team built a predictive model fusing claims, survey responses, and demographics to find the specific drug interactions and comorbidity patterns that precede falls. Outreach went from blanket coverage to targeted intervention, and the model correctly identified almost every actual fall patient before the event.
"The radiologist's eye still calls the verdict — but the model finds the lesion before the eye gets there."
Radiologists were reading chest X-rays at volumes that left no room for second-pass scrutiny — the model needed to do the looking that the human couldn't. The team trained a segmentation model to mark cancerous lung regions pixel by pixel, validated it iteratively against held-out test data, and integrated the predicted mask directly into the radiologist's workflow. The radiologist still makes the call — but now they're calling on a pre-marked image, not a blank one.
"Allocating savings used to be a quarterly negotiation — now it's a daily dashboard, and every initiative knows what it earned."
A national health insurer was running multiple cost-control initiatives simultaneously, with no fair way to attribute savings to each one. Budget season turned into a political negotiation. The team built an end-to-end framework — three years of claims, case management, and prior authorization data feeding budget models, savings estimation, FWA detection, and a fair-allocation engine. Each initiative now sees its actual contribution, in production, daily — and the dashboards work for C-suite, ops, and regulators alike.
"Refiling claims used to be a department — now it's a dashboard, and the denials we can avoid never get filed in the first place."
Claims refiling was eating headcount — a large team spent days reworking denials that were predictable from the moment the claim was generated. The team built a model that scores every claim on its way out the door and surfaces the underlying drivers. Primary payer and policy type turned out to be the dominant denial signals — actionable intelligence that fed renegotiation strategies and turned refile from a permanent department into a dashboard.
"Underpayment used to hide inside the aggregate — now every provider sees which payer underpaid which CPT code, and acts on it."
Providers knew they were being underpaid — they just couldn't see where. The aggregate numbers looked fine; the leakage was distributed across hundreds of payer-code combinations. The team built a clustering engine that mines the claim payment data across multiple dimensions and surfaces underpayments per payer and per denial reason — measured in denied dollars per invoice and ready for direct negotiation.
"The chart finally reads itself the way a physician would — into the right pathway, the right category, the right next step."
Clinical notes were piling up in inboxes that nobody had time to sort — every chart had to be read and triaged into a pathway by hand. The team fine-tuned a language model to read notes and route them into five disease categories with strong accuracy, freeing physicians to act on the chart instead of sorting it. The model is interpretable enough to debug, stable enough to deploy at scale, and reduces routing latency across the entire diagnostic queue.
"Discharge planning stopped beginning at day five — it now begins when the patient walks through the door."
Diabetic inpatient stays were stretching well past clinical necessity — and the care team only knew which patients would stay long after they already had. The team trained a neural network to flag high-LOS patients at the moment of admission, using demographics, comorbidities, and lab values. A parallel regression identifies the specific comorbidities driving extended stays — so the care team can plan discharge, intervene early, and forecast bed utilization before complications cascade.
"Our pediatric dentists stopped squinting at small mouths on small films — the model points where to look, and the chair time gets returned to the child."
Pediatric dental X-rays are hard — small mouths, restless patients, films that don't sit still. Dentists were spending more time finding the lesion than treating it. The team trained a detection model to identify caries and apical periodontitis automatically and overlay color-coded bounding boxes on the X-ray. The model points the dentist's eye to the lesion immediately — returning chair time to the patient and reducing the cognitive load of every read.
"We finally had a CHF readmission model that beat the public benchmarks — and the readmissions we prevented paid for the program in the first year."
CHF readmissions were the costliest preventable problem on the inpatient floor — and the standard CMS risk models weren't accurate enough to justify intervention spend. The team built a custom readmission model trained on 12 months of pre-admission history and 30 days of post-discharge data, and it outperformed both the CMS HF and CMS AMI benchmarks. The intervention engine layered on top now feeds tailored recommendations to discharge planners — and the readmissions prevented paid back the program in year one.
"Nephrologists got a real-time map of which patients were sliding toward dialysis — and intervened where the slide could still be slowed."
CKD Stage 3 patients progress toward dialysis at very different rates — and nephrologists were having to read every chart by hand to figure out who needed urgent intervention. The team trained a model on three to five years of longitudinal lab data and built dashboards that score patients by progression speed in real time. Nephrologists now see, at a glance, which patients are sliding fast — and intervene where the slide can still be slowed.
"Therapists used to triage by intuition — now the alert tells them which patient is hours away from an avoidable hospitalization."
A large home health and hospice provider was running on therapist intuition — visits prioritized by gut feel, with no early signal for which patients were sliding toward avoidable hospitalization. The team built a four-model platform on a unified Azure research lake — predicting acute hospitalization risk, 90-day mortality, ADL improvement, and mobility outcomes — and pushed the alerts directly to therapists in persona-tuned dashboards. Triage stopped being a guess.
"Pricing used to come from an actuarial table — now it comes from the data the actuary trusted but couldn't fully read."
Insurance charge models were stuck in traditional actuarial tables — useful, but not granular enough to support modern pricing decisions. The team built a regression model that learns charge patterns from individual risk factors — age, BMI, smoking, dependents, region — and surfaces feature importance transparently. Pricing teams now see why a given premium falls where it does, which segments to target, and where the actuarial table was leaving money on the table.
"Treatment planning used to start with a tracing pencil — now it starts with a mask, and the orthodontist gets back the half-hour."
Orthodontic treatment planning was starting with a tracing pencil — manual tooth segmentation on every panoramic X-ray, billable hours spent on a task no orthodontist enjoyed. The team trained a segmentation model to outline each tooth automatically with clinical-grade accuracy. The model integrates straight into the imaging workflow, returning half-hour blocks of time to the clinician — and the segmentation is consistent across the practice, not artist-by-artist.
"Pediatric films are harder, the patients more variable — the model still finds the structure, even when the mouth doesn't sit still."
Pediatric tooth segmentation is harder than adult — developing dentition varies more, films are noisier, and patients move. The team adapted a segmentation model for pediatric X-rays with carefully standardized preprocessing, and it now segments individual teeth reliably even on the difficult films. Orthodontists and pediatric dentists use the output for treatment planning, growth monitoring, and faster diagnostic workflows on the population that needs them most.
"Cross-sell stopped being a phone-list and started being a targeted conversation — fewer calls, higher conversion, less noise for the policyholder."
Insurance cross-sell was being run as a volume play — call every policyholder, hope a fraction convert. Conversion rates were low, customer fatigue was high. The team built a propensity model that scores health-insurance policyholders for vehicle-insurance interest using their existing data — demographic, policy, and vehicle features. Targeted lists replaced bulk calling, conversion climbed, and policyholders stopped getting calls they didn't want.
"Lesion borders used to depend on which dermatologist drew them — now the mask is the same on Monday morning as it is on Friday evening."
Dermatologists were drawing lesion borders by hand — and the borders moved depending on who drew them, when, and how tired they were. The team trained a segmentation model on standardized dermoscopic images with rotation and flip augmentation, and the model now outlines lesions pixel by pixel with consistent precision. Dermatologists use the mask as a baseline for screening — the consistency means earlier detection and fewer missed margins.
"Stroke risk used to live in the GP's instinct — now it lives in a number the patient can see and the population health team can act on."
Stroke risk assessment was being done in the GP's head — different doctor, different read. Population health teams had no consistent signal to organize prevention around. The team trained a classifier on clinical and lifestyle data — age, BMI, smoking, hypertension, glucose — and tuned it carefully to avoid bias toward majority classes. The model now produces a defensible stroke-risk score the patient can see and the population health team can act on at scale.
"Radiology second-reads used to depend on bandwidth — now the model gives every image a second look, and the senior radiologist arbitrates only the equivocal ones."
Senior radiologist second-reads were the gold standard for breast cancer screening — and the bottleneck. Junior reads went unverified, and the senior team couldn't keep up with volume. The team built a classifier that gives every image a second read automatically — benign or malignant — with augmentation and early-stopping tuned for clinical-grade generalization. Senior radiologists now spend their attention only on the equivocal cases, and the easy ones come pre-validated.
"A scan that took two senior radiologists to call now lands on the workstation with a probability and a Hounsfield trace — and the gut check happens faster."
CT classification for lung tumors had been a yes-or-no call — and that wasn't granular enough for treatment planning. The team trained a 3D model to classify scans into three categories — normal, malignant, benign — with preprocessing that normalizes Hounsfield Units and isolates the region of interest before classification. Radiologists now see a probability across three classes instead of a binary flag, and the model converges quickly and stably enough for reliable production support.