When a Hospital Network Stopped Making Clinical Decisions in the Dark
A large healthcare network was sitting on 12 years of patient data locked across 12 isolated EHR systems. Readmissions were avoidable. Deteriorations were predictable. We built a real-time analytics platform that finally connected the dots—improving patient outcomes by 28% in the first year.
Better patient outcomes
composite score across readmissions, complications, and recovery
Annual cost savings
driven by reduced preventable readmissions and optimised bed utilisation
Query response time
searching 12 years of data across 12 EHRs at population scale
Who We Worked With
Our client was a regional healthcare network spanning 8 hospitals and 14 outpatient facilities. While clinically respected, their technology had evolved in silos. Each facility ran its own EHR system—including two legacy platforms whose original vendors no longer existed. Twelve years of patient history was scattered. A patient seen at three different facilities had three separate records that no clinician could view simultaneously without requesting manual extracts.
"We had a patient readmitted four times in eight months. Each admission was treated as if it were the first. Different doctors, different facilities, same underlying condition that nobody connected because the dots were in four different systems. We had the data to flag her. We just couldn't see it."
— Chief Medical Officer
Company
Large integrated healthcare network
Team
3,800 clinical and administrative staff across 22 sites
Reach
Regional network spanning urban tertiary centres and rural district facilities
Platforms
Twelve years of clinical data, completely invisible to the clinicians who needed it most.
In a network of this scale, healthcare data isn't scarce—it’s fragmented. The data wasn't missing; it was simply trapped in 12 systems that had never been asked to speak to each other, creating blind spots at the exact moment clinical decisions were being made.
Predictable readmissions slipping through the cracks
An audit revealed 34% of 30-day readmissions shared visible risk factors at discharge (elevated markers, missing follow-ups). But pulling data from four different screens during a rushed 10-minute discharge workflow was impossible.
Bed management running on whiteboards and phone tag
Finding a transfer bed at the tertiary centre meant admissions teams calling each other for hours. Patients waited in emergency departments for beds that were actually empty, but digitally invisible.
Reactive, delayed outbreak responses
Seasonal disease patterns were tracked anecdotally. By the time a spike in respiratory or waterborne illnesses became obvious to the staff, the outbreak had already been developing for 10–14 days.
Clinical histories that took days to retrieve
When a physician needed records from a patient’s district-level visit six months prior, the manual request took 2–3 days. Doctors were forced to make immediate medication decisions based solely on the current admission.
Executives navigating by the rearview mirror
Monthly performance reports took the informatics team 12 days to manually compile. By the time the COO read them, the operational data was 6 weeks old.
They had tried building a centralized SQL data warehouse, but it was too technical for clinicians to use. They later added BI dashboards, but those only answered pre-configured executive questions. If a doctor had a specific clinical query, they still faced a 5-day wait for a manual report.
The CMO didn't need a sales pitch on big data; she needed proof. She carried the story of the patient readmitted four times as a systemic failure. The question wasn't if analytics mattered, but whether a 22-site organization could actually deploy them without the effort collapsing under its own weight.
We started with the decisions, not the dashboards.
Rather than starting with an abstract data architecture, we spent two weeks asking clinicians exactly what decisions they were forced to make with incomplete information. We worked backward from the workflow.
The data was not the bottleneck. The distance between data and decision was.
A dashboard a clinician has to log into during a chaotic discharge is useless. But a targeted alert that surfaces directly inside their existing workflow, calculating specific risk factors and suggesting an action? That changes behavior. We designed the platform around the decision, not the data.
We interviewed 42 stakeholders across 8 departments. The consensus was clear: the data existed, but it was fundamentally disconnected from the point of care. An audit of their 12 EHRs revealed a sobering reality—getting the data clean enough to trust would require serious engineering, not just a slick UI.
- Interviews with 42 emergency physicians, ward clinicians, and bed managers
- Technical audit of data models and APIs across all 12 EHR systems
- Workflow mapping for 8 high-impact clinical decision types
- 3-day shadowing of bed management and transfer operations
- Readmission cohort analysis validating the 34% preventable rate
Three non-negotiable rules: 1) Zero workflow disruption—integrate into existing screens. 2) No 'black box' AI—every recommendation must explain its reasoning in clinical terms. 3) Honest data—if data quality is poor, the UI must explicitly show the clinician it's poor so they can contextualize the risk.
- No rip-and-replace: We had to ingest data from all 12 legacy EHRs as they stood.
- Zero learning curve: The UI had to be intuitive enough for exhausted staff to use without training.
- Strict compliance: Full adherence to Indian health data regulations and HIPAA-equivalent audit logging.
- Self-sustaining: The client’s 6-person informatics team had to be able to maintain the system post-launch.
A clinical intelligence platform built to close the gap between insight and action.
We built six interconnected modules on a unified data architecture, translating 12 years of fragmented history into real-time, actionable clinical guidance.
Unified Health Data Lake
Ingests, normalises, and deduplicates patient data from all 12 EHRs in near real-time. It uses fuzzy matching (MRN, Aadhaar, demographics) to stitch together a single, longitudinal patient record.
Readmission Risk Engine
Calculates a predictive readmission score at the moment of discharge. High-risk patients trigger an alert directly inside the EHR, accompanied by suggested interventions (e.g., missing medication reconciliation).
Population Health Surveillance
Tracks disease presentation patterns geographically. If respiratory illnesses spike abnormally in a specific district, it alerts the public health team immediately.
Real-Time Bed Intelligence
A live, network-wide view of bed occupancy, filtered by speciality and dependency level, updated within 30 seconds of an admission or discharge.
Clinical Decision Support Alerts
Surfaces targeted warnings (drug interactions, deterioration risks, care gaps) inside the EHR based on the patient’s unified cross-facility history.
Executive Analytics Suite
Replaces the 6-week-old monthly reports with a live dashboard tracking occupancy, readmissions, and clinical outcomes, drillable down to the ward level.
Tech Stack
Apache Kafka & Spark
High-throughput streaming and historical data processing
Apache Druid
Sub-second query response at population scale for executive dashboards
FHIR R4 (HL7)
Standardized clinical data model unifying 12 disparate databases
LightGBM + SHAP
Predictive modeling with built-in, clinician-friendly explainability
React + Next.js
Fast, responsive frontends for dashboards and clinical portals
AWS (EMR, EKS, S3)
HIPAA-compliant, elastic infrastructure for data lakes and microservices
Risk scores explain 'Why'.
A '78% risk score' doesn't tell a doctor what to do. '78% risk because of 2 recent readmissions and no scheduled follow-up' does. Explainability isn't a buzzword; it's the only way to make AI clinically actionable.
Visible data quality scores.
If a patient's allergy history is only 60% complete, the UI explicitly states '60% Confidence'. Clinicians distrust platforms that pretend to know everything. Visible uncertainty actually builds trust.
Twenty weeks to launch. And a massive data cleanup that was worth every delayed day.
We planned for messy data, but reality was worse. We intentionally delayed the clinical launch by three weeks just to remediate legacy EHR data. In healthcare, shipping on poor data isn't a bug—it’s a patient safety risk.
Discovery & Architecture
Weeks 1–7Workflow mapping, Kafka pipeline build, and API adapters developed for all 12 systems. Initial profiling revealed severe data inconsistencies in 4 legacy systems.
Data Quality Remediation
Weeks 8–11A heavy three-week sprint dedicated purely to cleaning duplicates, non-standard medications, and missing values. The CMO backed the delay without hesitation to ensure clinical safety.
Models & Population Surveillance
Weeks 12–15Readmission and deterioration models trained and validated against hold-out datasets. Surveillance algorithms calibrated against 18 months of historical outbreak data.
Bed Intelligence & EHR Integration
Weeks 16–18Real-time occupancy network deployed. Clinical governance committee pressure-tested and approved 24 distinct alert types before EHR integration.
Full Rollout & Knowledge Transfer
Weeks 19–20Phased launch across 22 sites. We spent two full weeks pair-working with the internal informatics team to ensure they could independently manage and scale the platform.
- 1 × Engagement Lead
- 2 × Data Engineers (Kafka, Spark, FHIR)
- 1 × ML Engineer (Readmission & Surveillance Models)
- 2 × Backend Engineers (Alerts & Bed Intelligence)
- 1 × Frontend Developer
- 1 × Product Designer
Our monthly clinical governance reviews were the battlefield where alert logic was forged. Physicians aggressively challenged model outputs, negotiating the exact boundary between what the system should recommend and what should be left to clinical judgment. That friction is what made the system safe.
Challenges & How We Solved Them
The Challenge
Two legacy EHRs were so old the original vendors were out of business. One was a 15-year-old undocumented database built by a retired sysadmin.
How We Handled It
We located the retired sysadmin and brought him in for two days to help reverse-engineer the schema. We built read-only adapters to protect the fragile environments and safely migrated 80,000 patient records with zero data loss.
The Challenge
The readmission risk model performed poorly on maternity patients, as the clinical predictors for obstetrics differ wildly from general medicine.
How We Handled It
We built a separate, specialized obstetric model. Because the data pool was smaller, we extended the training window to 36 months and explicitly communicated the model's evolving confidence levels to the maternity ward. Honesty won their buy-in.
The Challenge
Initial alert fatigue. In the first two weeks, doctors were blindly ignoring two specific alert types, dropping acceptance rates below 30%.
How We Handled It
We didn't blame the users; we blamed the design. We suspended the alerts, narrowed the trigger conditions with the clinical leads to remove false positives, and relaunched. Acceptance rates immediately jumped to 75%.
Twelve months later: 28% better outcomes, real-time beds, and proactive decisions.
The metrics were incredible, but the cultural shift was profound. Physicians now openly discuss AI risk scores during discharge rounds. Bed managers orchestrate transfers with total visibility. The data finally became part of how the network thinks.
28%
Better patient outcomes
composite improvement across readmissions, complications, and recovery
₹1.9Cr
Annual cost savings
driven by optimized bed usage and plummeting readmission rates
35%
Admin time saved
freed informatics staff from manual reporting; freed doctors from records hunting
12+
EHRs unified
flawless integration across 22 sites with zero data loss
99.9%
Platform uptime
zero workflow-impacting outages in year one
< 1 sec
Query response
sub-second analytical lookups on 12 years of history
The surveillance system caught an atypical pneumonia cluster 9 days before manual tracking would have, allowing the public health team to mobilize resources and alter the outcome for at least 40 patients.
The 12-day executive reporting cycle was eliminated. The COO now manages network capacity daily based on real-time realities, not historical anecdotes.
Readmissions for high-risk chronic patients dropped from 18.4% to 11.2%. The discharge team attributed this directly to the in-workflow EHR alerts prompting mandatory medication checks.
"Before this platform, I walked into morning bed meetings working off a printout from the night before, while my managers relied on hours-old phone calls. Now, the screen shows me exactly where we are—network-wide, right now. Last Tuesday we spotted a capacity crisis 14 hours before it hit and pre-positioned staff to handle it. That doesn't make the news. But that's what this platform does for us, every week."
Chief Operating Officer
Healthcare Network Client
Learnings That Outlasted the Project
Data quality is a clinical safety issue, not an IT nuisance.
You can't 'fix it in post' when it comes to healthcare data. A readmission model fed on bad data produces dangerous recommendations. Delaying our launch to scrub 12 years of legacy records is the exact reason clinicians trusted the platform on day one.
Alert fatigue is a UX failure.
When doctors ignore alerts, it means the system is noisy and irrelevant. Treating low acceptance rates as a design flaw—and actively tuning the logic in clinical governance meetings—is mandatory for adoption.
The distance between data and decision is everything.
A beautiful dashboard hidden behind a login screen will never be used during a chaotic shift. Analytics only matter when they are injected directly into the moment the decision is being made.
Related Work & Services
Running a healthcare network where your data exists, but your decisions don't benefit from it? That gap is exactly what we're built to close.
Services Used in This Project
Your clinical data already knows things your team doesn't. Can it tell them in time?
We build analytics platforms for networks with deeply fragmented systems, legacy tech, and exhausted staff. We know what it takes to go from messy data to clinical decision support that actually gets used. Tell us about your data landscape, and we'll give you an honest view of what's achievable.
No generic big data pitch. A real conversation about your clinical reality.