Flagged

Maya Krishnan stared at the screen as the dashboard refreshed, the rows of numbers scrolling past like a tide of quiet bureaucracy. She had been with Arogya Health for six years, the sort of tenure that made her name appear on the internal distribution list for any change in the claims processing engine. Today, her task was the quarterly audit of the “high-risk cohort”—a list of policyholders whose projected cost to the insurer exceeded a threshold determined by a logistic regression model that had been trained on three years of claims data, demographic variables, and a smattering of wearable-device metrics.

She clicked open the CSV export, the columns aligning neatly: “MemberID”, “Age”, “BMI”, “ChronicScore”, “RiskPercentile”. The spreadsheet was a familiar landscape, the kind of thing she could read blindfolded after a while. The cohort usually contained a few hundred names, each a datapoint that her team would flag for closer review, maybe a phone call from a case manager, or a prompt to adjust the member’s premium. The risk model was a black box to most, but to Maya it was a set of rules she could trace back to a Jupyter notebook she’d once written to debug an outlier.

She filtered the list for the city code “DEL”, then sorted by “RiskPercentile” descending. The top ten names were of people she recognized from previous audits—one who had been hospitalized for a triple bypass, another who’d filed a claim for a rare autoimmune disorder. Maya’s eyes scanned the list, noting the usual suspects, when a name appeared that made her pause: “Maya Krishnan, 32, 23.7, 4.1, 97.3”.

She blinked, then read the line again. The entry was an exact match for her own member ID, which she knew because she had once requested a copy of her policy for a personal budgeting spreadsheet. The risk percentile was 97.3, placing her in the top three percent of projected cost to Arogya Health. The model had flagged her as a likely high-cost claimant, eleven days before the night she would later spend in a hospital bed with a ruptured aneurysm that no one else would ever know she feared.

Maya’s hand hovered over the mouse. She could have dismissed it as a glitch—perhaps a data entry error, a misaligned row. She had seen similar mistakes before, a stray comma here, a duplicate ID there. But the log timestamps were clean, the data pull had been verified by the data-engineering team an hour ago. She opened the member profile, a compact pane that displayed the raw inputs the algorithm used: age, BMI, last three months of claim amounts, a “SleepScore” derived from a wearable, and a “StressMetric” that Arogya collected from an optional mobile app. The “SleepScore” was 38, far below the healthy threshold of 70; the “StressMetric” was 89, which the model treated as a strong predictor of future hospitalization.

She scrolled down to the “Feature Contributions” section, a table that broke down the weight each variable had on the final risk score. “Age” contributed +0.12, “BMI” +0.05, “SleepScore” –0.18, “StressMetric” +0.31, “RecentClaims” +0.24. The sum of the weighted features was the log-odds that the model transformed into a probability. The math was simple enough: logit(p) = Σw_i x_i, p = 1 / (1 + e^–logit). Maya could see the numbers, she could even recompute the probability on a scrap of paper if she wanted.

She felt a flicker of unease.

She leaned back, her chair creaking under the shift, and let the hum of the office wash over her. The cubicle walls were plastered with posters about “Data-Driven Care” and “Predictive Analytics Saves Lives”. A coffee mug sat on the desk, half-filled with cold brew, a testament to the long hours the analytics team often pulled. Outside, the Delhi monsoon had begun, a thin curtain of rain slashing against the glass, making the city look like a circuit board seen through a fog.

Maya opened the audit log to see who had last touched the cohort file. It showed that the model had run at 02:15 GMT, and the risk scores had been written to the database at 02:37 GMT by the ETL pipeline. No human had edited the file since. The flag was entirely algorithmic.

She typed a quick query into the internal ticketing system, opening a ticket to the “Model Governance” team. “Flagged cohort entry for MemberID 84219 (Maya Krishnan). Need verification of data integrity and feature attribution.” She attached the screenshot, the CSV slice, and a note that was more a question to herself than to anyone else: “Why am I in this cohort?”

She expected an automated reply, perhaps an acknowledgement from the bot that managed the ticket queue. Instead, after a few minutes, a message pinged from Arjun, her lead on the claims analytics team. “Maya, you’re in the high-risk list? That can’t be right. Let’s pull the raw logs and see if the ID got mixed up.”

Maya smiled despite the knot tightening in her chest. “Sure, Arjun. I’ll send you the export I have.”

Arjun’s reply was brisk: “Good catch. I’ll look into it. This is probably a data hygiene issue.” He added a winking emoticon, the kind that made the corporate chat feel marginally less sterile.

She closed the ticket and opened a fresh notebook, importing the model weights from the repository. The code was clean, version-controlled, with a README that explained each feature’s source. The “StressMetric” was a composite of heart-rate variability and a self-reported questionnaire, both of which Maya had never opted into. The “SleepScore” came from a third-party app she’d tried for a week a year ago and then abandoned. The data pipeline, according to the comments, defaulted missing entries to the median of the cohort, which for “StressMetric” was 45 and for “SleepScore” was 71. But in her profile, both values were present, suggesting she must have inadvertently enabled the data sharing.

Maya clicked through the API logs, a series of JSON payloads that recorded each data ingestion event. The timestamps aligned with the dates she had remembered loading the app on her phone, the evenings she had ignored the notification to sync her wearable. She could see the “DeviceID” hash, the “UserConsent” flag set to true, and the “DataPush” status marked “success”. The system had dutifully collected the data, even though she had never consulted it.

She leaned forward, her elbows resting on the desk, and thought about the model’s perspective. To the algorithm, she was not a claims analyst with a secure job and a stable income; she was a vector of numbers, each weighted by a coefficient derived from a training set that never knew her morning commute, the fact that she still bought chai from a roadside stall, or the occasional night she stayed up to listen to a friend’s heartbreak over a cracked speaker. The model had no concept of intent, only correlation. The fact that her “StressMetric” was high could be due to a temporary bout of insomnia, a project deadline, or an early symptom of something more serious—something she had not yet told anyone.

A soft chime indicated a new email. It was from the “Data Integrity” team, headed by Priya, who was known for her meticulousness and for never using the word “sorry.” The subject line read: “Re: Flagged Cohort Entry – MemberID 84219”. The body was brief: “Maya, we’ve traced the entry to the data ingestion pipeline. All fields are present and validated. No duplication detected. The risk score is consistent with the current model version (v3.2.1). Let us know if you need further assistance.”

Maya stared at the words. The algorithm had not only flagged her; it had done so with confidence, using data she had never consciously given. The mundane horror of the discovery was that the system knew more about her hidden moments than she did.

She decided to test the model. She opened a sandbox environment, fed a synthetic profile identical to hers except for the “StressMetric”, setting it to the cohort median of 45. The risk percentile dropped to 62.3, well below the high-risk threshold. She restored the original value, and the percentile surged back to 97.3. The model’s sensitivity to that single feature was stark.

She wrote a short note in the ticket, attaching the sandbox results: “Changing StressMetric from 89 to 45 reduces risk percentile from 97.3 to 62.3. Indicates heavy weighting on this feature. Question: why is this metric present if user did not opt-in?”

She sent it, then leaned back again, the rain now a steady drum against the glass. The office lights were dimming as the day shifted toward evening, the fluorescent tubes flickering like a dying star. Maya’s mind drifted to the night eleven days from now—a night she had marked on her personal calendar with a red circle, a reminder to call her mother about the upcoming family wedding. She had been feeling a dull ache behind her eyes for a week, dismissing it as fatigue from the new project at work. She had not mentioned it to anyone, not even Arjun, not even Priya. The model, however, had already flagged her.

A knock on the cubicle wall announced the arrival of Raj, the facilities manager, holding a tray of fresh samosas. “Hey Maya, you look like you’ve seen a ghost. Want one?”

Maya forced a laugh. “Sure, why not. I’m just… surprised the algorithm knows more about me than I do.”

Raj raised an eyebrow, the corners of his mouth twitching. “Those things are getting smarter. I heard they can predict if you’ll get a cold before you sneeze.”

Maya took a samosa, the flaky pastry warm against her palm. “If they can predict a cold, I suppose they could predict a heart attack too,” she said, the dry humour surfacing despite the rising tide of unease.

He chuckled, a short burst that echoed off the partition. “Just don’t let them predict your salary. That’d be a disaster.”

She smiled, the tension easing just enough to make the moment bearable. She placed the samosa on her desk, the oil glistening under the fluorescent light. The bite was crunchy, the filling spiced, a small comfort in a sea of numbers.

She opened a new document, the title a single line: “Personal Log – 07/12/2026”. She wrote: “The model flagged me eleven days before the episode. I am a data point. I am also a person who enjoys samosas on rainy days.”

She saved the file, then closed the laptop with a soft click. The room hummed with the low whirr of servers in the data center behind the glass wall, the unseen heart of Arogya Health beating in rhythm with the city’s monsoon. Maya gathered her things, the weight of the discovery settling like a quiet pressure in her chest. She walked toward the exit, the rain blurring the world outside, each drop a reminder that even the most sophisticated algorithm could not account for the simple fact that she was still, fundamentally, human.