The Chart the Engine Couldn't Read
On her last week as a medical coder, Marian finds a chart the autonomous system routes to her queue — and discovers why, after twenty-three years, the machine still occasionally needs a human to read what clinicians actually wrote.
The Chart the Engine Couldn't Read
The queue had thirty-two charts in it when Marian logged in Monday morning. By lunchtime, four.
This was how it worked now. The engine — nobody at Parkstream Health System called it by its vendor name anymore, the way nobody had called the autonomous vehicle a Waymo for years, it was just the engine — processed the ED charts through the night. Anything it coded above 0.93 confidence went straight to billing. Anything at or below 0.93 came to the human queue. Marian's queue. Or what was left of it.
She had been a certified medical coder for twenty-three years. She had come in through Parkview Community College's Health Information Technology program in 2003, back when ICD-9 was still the coin of the realm and everyone in her cohort thought ICD-10's 2015 rollout was going to be the biggest change they would ever see in their careers. It turned out to be the second biggest.
The biggest was this. Four charts. A Monday morning queue of four charts.
She worked through them in the order the engine had flagged them — which was itself a feature Nym had rolled out last summer, a little ranking system that told her which charts the model was most uncertain about. She appreciated the honesty. It was better than pretending.
The first was a lacerated hand from a Saturday night ED visit. The engine had coded it correctly, actually; Marian could see that at a glance. It had just flagged itself because the operative note used an unusual phrasing for the suture technique. Five minutes. Approve. Bill.
The second was a pediatric asthma exacerbation where the documentation mentioned a home nebulizer history the engine wasn't sure how to weigh in the severity assessment. She looked at the nursing notes, looked at the attending's impression, adjusted the ICD-10-CM code from J45.901 to J45.41, and sent it. Eight minutes.
The third was a straightforward abdominal pain workup. The engine was uncertain because the final diagnosis in the attending's note was resolved pain, etiology unclear, likely functional, and the engine had been trained not to code diagnoses that used hedging language. Marian coded R10.9 and moved on. Four minutes.
The fourth was the one.
The fourth was a seventy-eight-year-old woman who had come in at 11:42 PM on Saturday night with her husband. The chief complaint was logged as "family concerned about patient behavior." The triage note said the husband had driven her in because she had become increasingly confused over the past week and that morning she had been unable to name their daughter.
Marian sat back.
The engine had routed the chart to her with a confidence of 0.62, which was low. The note said flagged for human review: multiple differential diagnoses, documentation inconsistency, high-stakes coding implication.
She opened the full chart.
The attending's H&P was thorough. Progressive cognitive decline over several months, accelerated over the past week. The husband reported wandering at night. The patient had been diagnosed with mild cognitive impairment eighteen months prior at a memory clinic. Labs were unremarkable. CT head showed age-related atrophy, no acute findings. The attending had ordered a UA, which came back with trace leukocytes and nitrites. The patient had been admitted for evaluation of acute-on-chronic encephalopathy, rule out urinary tract infection as precipitant.
The engine's candidate codes were:
- R41.0 (disorientation, unspecified)
- N39.0 (urinary tract infection, site not specified)
- F03.90 (unspecified dementia without behavioral disturbance)
The engine's uncertainty was about sequencing — which code to put first, the dementia or the delirium or the UTI — and about whether to code the dementia at all when the ED attending had not personally diagnosed it.
Marian read the chart twice more.
Then she opened the consultant notes.
The ED attending had called a geriatrics consult at 4 AM. The consultant had documented a careful history with the husband and concluded — and Marian read this sentence three times before she understood it — that the precipitating factor for the acute confusion was not the UTI but a recent change in the patient's Aricept dosing, made at a memory clinic visit eleven days earlier, which had not been communicated to the ED or the patient's primary care physician.
The chart was not a UTI with delirium superimposed on dementia. The chart was a medication-induced delirium misidentified as an infection-related delirium because an outpatient dosing change had never made it into the EHR.
Marian sat at her desk and looked at that for a long time.
The engine had never read the consultant note. It had read the attending's admission assessment, the labs, the CT results, and the nursing notes. It had not read the 6 AM geriatrics consult that had been uploaded into the chart as a scanned PDF from an outside clinician's letterhead because Parkstream's Epic instance and the memory clinic's were on different HL7 versions and the integration project had been delayed since February.
The engine was not wrong. The engine was reading exactly the documentation it was designed to read. The engine just didn't know the consultant note was there.
Marian opened the encounter-level sequencing tool, added T43.205A (adverse effect of unspecified antidepressants and antidementia drugs, initial encounter) as the primary, F05 (delirium due to known physiological condition) as the secondary, and F03.90 (unspecified dementia) as tertiary. She dropped the N39.0. She wrote a free-text coding query to the attending asking for documentation confirmation of the medication adverse effect etiology.
The billing impact of her code change was negligible — maybe $180 on the DRG weight, she could never remember the exact weights without looking them up.
The clinical impact was that the patient's chart would now correctly reflect what the geriatrics consultant had determined, which meant the primary care physician who saw her at follow-up would see medication adverse effect in the patient's problem list instead of UTI and would know to reconcile her medications. Which might prevent this from happening again.
Marian clicked submit.
The engine logged her override as training data. Next time — next week, next month, Nym's release cycle was quick now — the engine would look at scanned consultant notes differently. It would pull the text through OCR. It would check for medication attribution language. It would learn.
She knew this. She had been to the vendor training last October. She had watched the product manager, a very earnest young man who had apparently never worked a day in a coding department, explain that the engine improved with every human override and that this was why the "human in the loop" remained critical to the system's performance.
She had wanted to ask him what happened when there were no more humans to be in the loop. She had not asked, because the answer was obvious and it would have embarrassed him.
At four that afternoon her queue was empty and she sat with a cup of coffee and watched the sunlight come sideways through the window behind her monitor.
Her supervisor, Yvette, appeared in the doorway. Yvette had been a coder herself once, twenty-five years ago, and was now a revenue cycle director with seventy-one direct and indirect reports in 2022 and fourteen in 2026.
"You got a minute?" Yvette asked.
Marian nodded.
"I wanted to tell you before the email went out tomorrow. Your last day is the fifteenth."
"Okay."
"The severance is ten months. HR will send the paperwork tomorrow. There's a coding audit position open at Bayshore Regional if you want it, it's a step down in pay but it's there. I already sent them your name."
"Thank you."
"The transition plan for the last three weeks —"
"I know, Yvette. I've been here twenty-three years. I know the plan."
Yvette looked at her for a moment. Marian watched her older supervisor's face soften in a way Marian had seen before, usually when Yvette was about to say something she had been rehearsing.
"You caught the medication adverse effect one this morning."
"Yeah."
"The engine flagged it to me too. It wanted me to know a coder had done something unusual."
Marian laughed, the first time she had laughed all day. "Of course it did."
"I looked at it. You sequenced it right. The engine didn't have the consultant note. The OCR pipeline doesn't pick up those scanned PDFs from the memory clinic yet." Yvette hesitated. "I wanted you to know that's the kind of thing the vendor is still six to nine months away from handling. Maybe longer. The integration gaps are not going away on the timeline they told us."
Marian nodded.
"I'm not saying you shouldn't take the severance. I'm saying I want you to know what you did this morning mattered."
"It mattered to one patient."
"Yeah."
"The engine will get to the scanned consultant notes eventually."
"Eventually, yes."
"And in the meantime there will be charts like this one that go through without anyone catching them."
Yvette did not answer that. She did not need to. They had both been doing this work long enough to know how it ended.
"I have fifteen more days," Marian said. "I'll work the queue. Tell the engine I said hello."
Yvette laughed in spite of herself and then went back down the hall.
Marian logged out at 4:47. She walked to her car in the employee lot behind the hospital. It was still light. The spring had been slow to arrive in Ohio this year and the trees at the edge of the lot were just starting to leaf out, green bursting through the gray branches in a way that always surprised her even after twenty-three years of watching it happen from this same parking lot.
She drove home. She made dinner for her husband, who had been a high-school shop teacher and had retired two years ago when the program was eliminated because nobody was enrolling in metalworking anymore because the graduates could not get jobs at the factories because the factories had machine operators who supervised twelve CNCs each, not six. He had taken the retirement package. He had a pension.
Marian had a 401(k) with thirty-one years of steady contributions and a ten-month severance and whatever COBRA would cost from May through December. She had one working hand and one slightly arthritic hand and an associate's degree from 2003 and an AAPC CPC credential that, she understood, would continue to exist but would not continue to be economically valuable the way it had been from 2003 to about 2024.
At dinner her husband asked her how her day had gone.
"Fine," she said. "I coded a chart the engine couldn't read."
He looked up from his plate. "One of the ones they pay you extra for?"
"No. Just a chart. A woman with a medication adverse effect that got missed because the consultant note was in a different file format."
"And the machine didn't catch it."
"Not yet."
"But you did."
"This time."
He nodded. They finished dinner. Afterward she went and sat on the back porch and watched the long spring dusk come down and thought about the woman with the medication adverse effect and whether her primary care physician, when she went back for follow-up, would reconcile her medications properly. She thought about all the other women like her, all over Ohio and Indiana and Kentucky and the entire country, whose charts were being coded tonight by engines that could not yet read the scanned consultant notes from memory clinics on different HL7 versions.
She thought about which woman in which hospital, on which night, would be the last one to have a person on the other end of the chart.
She did not know. The engines would get to the scanned notes eventually. Eventually, yes. In the meantime there would be a diminishing number of Marians sitting in a diminishing number of revenue cycle offices catching a diminishing number of edge cases until one day, quietly, with no ceremony and no press release, the last edge case would be caught too.
She watched the dusk. The trees were still putting out their spring leaves. She sat on the porch a long time before she went inside.
If this story moved you, you might also want to read the accompanying analysis of how autonomous medical coding eliminates 430,000 revenue cycle jobs by 2032, or the short story The Benchmark We Built — another quiet account of work changing.